As of: 2026-10-25 23:49 CET.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/topics/media-generation/

# Media generation: everything we have

42 events (27 after June 2026), 22 models, 13 videos, 20 posts. Latest event 19 October 2026.

Importance: * minor to ***** historic. Confidence says how well the event is sourced. Each event has its own page at https://postcutoff.com/e/<id>/.

## Key moments

- 2021-01-05: OpenAI unveils DALL·E and CLIP (****) https://postcutoff.com/e/2021-01-05-dall-e-clip/
- 2022-04-06: DALL·E 2 brings photorealistic text-to-image generation (****) https://postcutoff.com/e/2022-04-06-dall-e-2/
- 2024-02-15: OpenAI previews Sora, a text-to-video 'world simulator' (****) https://postcutoff.com/e/2024-02-15-sora/
- 2025-05-20: Google's Veo 3 generates video with native audio (****) https://postcutoff.com/e/2025-05-20-veo-3/
- 2025-09-30: OpenAI launches Sora 2 and the Sora social app (****) https://postcutoff.com/e/2025-09-30-sora-2/
- 2026-05-19: Google unveils Gemini Omni, an any-to-any model that generates and conversationally edits video (****) https://postcutoff.com/e/2026-05-19-gemini-omni/
- 2026-07-23: Black Forest Labs unveils FLUX 3: one model for images, 20-second video with audio, and robot actions (****) https://postcutoff.com/e/2026-07-23-black-forest-labs-flux-3/

## Timeline, oldest first

### 2021-01-05: OpenAI unveils DALL·E and CLIP

****; OpenAI; Media generation; confidence high; https://postcutoff.com/e/2021-01-05-dall-e-clip/

DALL·E generated images from text prompts using a 12B-parameter Transformer, and CLIP learned joint image–text representations from 400M image-caption pairs; CLIP became a key component of later diffusion image generators.

- Both announced 5 January 2021
- DALL·E: 12-billion-parameter version of GPT-3 trained on text–image pairs
- CLIP: trained on 400M image–text pairs; strong zero-shot ImageNet accuracy
- CLIP weights open-sourced; used by Stable Diffusion's text encoder (v1)

#### What happened
OpenAI introduced a text-to-image model and a contrastive vision-language model on the same day.

#### Why it matters
Launched the text-to-image era and made natural language the interface for vision models.

Sources: [DALL·E: Creating images from text (OpenAI)](https://openai.com/index/dall-e/); [CLIP: Connecting text and images (OpenAI)](https://openai.com/index/clip/); [Learning Transferable Visual Models From Natural Language Supervision (arXiv)](https://arxiv.org/abs/2103.00020); [Zero-Shot Text-to-Image Generation (arXiv)](https://arxiv.org/abs/2102.12092)

### 2022-04-06: DALL·E 2 brings photorealistic text-to-image generation

****; OpenAI; Media generation; confidence high; https://postcutoff.com/e/2022-04-06-dall-e-2/

OpenAI's DALL·E 2 used a diffusion decoder conditioned on CLIP embeddings to generate high-resolution, photorealistic images from text, kicking off 2022's image-generation boom alongside Midjourney and Stable Diffusion.

- Announced 6 April 2022
- Paper: 'Hierarchical Text-Conditional Image Generation with CLIP Latents', arXiv 2204.06125
- Supported inpainting and image variations
- Opened to the public without a waitlist in September 2022

#### What happened
DALL·E 2 produced images of a quality that made text-to-image a mainstream phenomenon.

#### Why it matters
Marked diffusion models' takeover of image generation and triggered debates on artists' rights and synthetic media.

Sources: [DALL·E 2 (OpenAI)](https://openai.com/index/dall-e-2/); [Hierarchical Text-Conditional Image Generation with CLIP Latents (arXiv)](https://arxiv.org/abs/2204.06125)

### 2024-02-15: OpenAI previews Sora, a text-to-video 'world simulator'

****; OpenAI; Media generation; confidence high; https://postcutoff.com/e/2024-02-15-sora/

OpenAI previewed Sora, a diffusion-transformer model generating up to a minute of high-fidelity video from text, framing video generation as a path toward general-purpose simulators of the physical world.

- Previewed 15 February 2024 (red-teamers and selected artists only)
- Generated videos up to one minute long
- Diffusion transformer operating on spacetime patches
- Publicly released to ChatGPT Plus/Pro users on 9 December 2024
- Succeeded by Sora 2 in September 2025

#### What happened
OpenAI released sample videos showing a dramatic leap in coherence, length and realism over previous text-to-video systems.

#### Why it matters
Reset expectations for AI video overnight and triggered a video-generation race (Veo, Kling, Runway Gen-3).

Sources: [Sora (OpenAI)](https://openai.com/index/sora/); [Video generation models as world simulators (OpenAI technical report)](https://openai.com/index/video-generation-models-as-world-simulators/)

### 2025-05-20: Google's Veo 3 generates video with native audio

****; Google DeepMind; Media generation; confidence medium; https://postcutoff.com/e/2025-05-20-veo-3/

Announced at Google I/O 2025, Veo 3 generated video with synchronized sound effects, ambient noise and dialogue from text prompts, producing clips that went viral for their realism.

- Announced 20 May 2025 at Google I/O
- Native audio generation including dialogue and lip sync
- Launched with Flow, an AI filmmaking tool
- Initially available to Google AI Ultra subscribers in the US

#### What happened
Google released a video model that produces sound and speech together with visuals.

#### Why it matters
Crossed the uncanny valley for short AI video with dialogue, intensifying concerns about synthetic media.

Sources: [Veo (Google DeepMind)](https://deepmind.google/models/veo/); [Wikipedia: Veo (text-to-video model)](https://en.wikipedia.org/wiki/Veo_(text-to-video_model))

### 2025-08-26: Google releases Gemini 2.5 Flash Image ('Nano Banana')

***; Google DeepMind; Media generation; confidence medium; https://postcutoff.com/e/2025-08-26-gemini-2-5-flash-image/

Google launched Gemini 2.5 Flash Image, nicknamed 'Nano Banana', an image generation and editing model notable for character consistency and conversational multi-turn editing, which drove a surge of Gemini app adoption.

- Released 26 August 2025
- Topped LMArena image-editing leaderboard under the codename 'nano-banana' before launch
- Strong character/subject consistency across edits
- Outputs carry SynthID invisible watermark

#### What happened
Google shipped an image model that made precise, prompt-based photo editing a viral consumer phenomenon.

#### Why it matters
Showed natively multimodal LLMs overtaking specialized diffusion tools for image editing.

Sources: [Introducing Gemini 2.5 Flash Image (Google Developers Blog)](https://developers.googleblog.com/en/introducing-gemini-2-5-flash-image/); [Image editing in Gemini just got a major upgrade (Google)](https://blog.google/products/gemini/updated-image-editing-model/); [Wikipedia: Nano Banana](https://en.wikipedia.org/wiki/Nano_Banana)

### 2025-09-30: OpenAI launches Sora 2 and the Sora social app

****; OpenAI; Media generation; confidence high; https://postcutoff.com/e/2025-09-30-sora-2/

OpenAI released Sora 2, a video-and-audio generation model with improved physical realism and synchronized dialogue, alongside an invite-only iOS social app featuring 'cameos' of users' own likeness; the app quickly reached #1 on the US App Store.

- Announced 30 September 2025
- Generates synchronized dialogue and sound effects
- Sora iOS app with 'cameos' (consented likeness insertion)
- Sparked copyright and likeness controversies in its first weeks

#### What happened
OpenAI paired a much-improved video model with a TikTok-style feed of AI-generated videos.

#### Why it matters
Turned AI video into a mass social medium and intensified debates about deepfakes, likeness rights and copyright.

Sources: [Sora 2 is here (OpenAI)](https://openai.com/index/sora-2/); [Sora 2 System Card (OpenAI)](https://openai.com/index/sora-2-system-card/)

### 2026-02-05: Kling 3.0: unified multimodal video model with native audio and multi-shot 'AI Director'

***; Kuaishou, Kling AI; Media generation; confidence medium; https://postcutoff.com/e/2026-02-05-kling-3-0/

Kuaishou launched Kling 3.0 on 2026-02-05, a rebuilt unified multimodal architecture that generates up to 15-second clips with native audio and lip-sync, and can compose up to 6 shots in one clip with automatic continuity.

- Release: 2026-02-05 (Kuaishou IR)
- Clip length up to 15 s (from 10 s), native multilingual audio and lip-sync
- Multi-shot 'AI Director': up to 6 shots per 15-second clip, each with its own framing and camera
- Third-party sources claim native 4K / 60 fps (unverified)

#### What happened
Kling 3.0 rebuilt the model as one multimodal system taking text, image, audio and video as inputs and outputs, adding shot-by-shot direction within a single generation.

#### Why it matters
It set the bar for Chinese video models in early 2026 and was followed by Kling 4.0 in September.

Sources: [Kuaishou IR: Kling AI launches 3.0 model](https://ir.kuaishou.com/news-releases/news-release-details/kling-ai-launches-30-model-ushering-era-where-everyone-can-be); [Kling 3.0 model page](https://kling.art/model)

### 2026-02-18: Google launches Lyria 3: song generation with vocals in the Gemini app

***; Google DeepMind, Google; Media generation; confidence high; https://postcutoff.com/e/2026-02-18-google-lyria-3-gemini-app/

Google put Lyria 3 into the Gemini app, letting adults generate 30-second songs with vocals and auto-written lyrics from text, photos or videos in 8 languages, all SynthID-watermarked; on 2026-03-25 Lyria 3 Pro added ~3-minute structured songs and developer access (Gemini API, Vertex AI).

- Gemini app: 30 s tracks with vocals + lyrics, Nano Banana cover art, 18+ only, higher limits for AI Plus/Pro/Ultra
- Languages: English, German, Spanish, French, Hindi, Japanese, Korean, Portuguese
- Gemini app can check uploaded audio for SynthID watermarks
- YouTube Dream Track (Shorts soundtracks) moved to Lyria 3
- 2026-03-25 Lyria 3 Pro: up to ~3 min with intro/verse/chorus/bridge control; API ids lyria-3-pro-preview ($0.08/song) and lyria-3-clip-preview ($0.04/clip)

#### What happened
Lyria 3 was Google's first Lyria model to sing: it generates vocals and writes lyrics, not just instrumentals like Lyria 2. It rolled out on desktop on 2026-02-18, then mobile. Prompts naming an artist are treated as broad inspiration, not imitation.

Five weeks later (2026-03-25) Google released Lyria 3 Pro for full ~3-minute songs with structure control and opened both models to developers in the Gemini API/AI Studio and in public preview on Vertex AI, plus Google Vids and ProducerAI. The Pro launch cited producer Yung Spielburg's Lyria-assisted score for the DeepMind short film "Dear Upstairs Neighbors".

#### Why it matters
It put a Suno-class song generator in front of Gemini's mass consumer audience and gave developers a first-party, pay-per-song music API with watermarking and C2PA credentials.

Sources: [Google: Use Lyria 3 to create music tracks in the Gemini app](https://blog.google/innovation-and-ai/products/gemini-app/lyria-3/); [Google: Lyria 3 expands to more Google products (Lyria 3 Pro)](https://blog.google/innovation-and-ai/technology/ai/lyria-3-pro/); [Workspace Updates: custom soundtracks with Lyria 3](https://workspaceupdates.googleblog.com/2026/02/create-custom-soundtracks-with-lyria-3.html); [Vertex AI Lyria 3 model page](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/lyria/lyria-3); [Music Business Worldwide on Lyria 3](https://www.musicbusinessworldwide.com/google-just-launched-lyria-3-its-most-advanced-ai-music-generator-yet-in-the-gemini-app/)

### 2026-02-26: Google launches Nano Banana 2 (Gemini 3.1 Flash Image)

***; Google DeepMind, Google; Media generation; confidence high; https://postcutoff.com/e/2026-02-26-nano-banana-2/

Nano Banana 2 — technically Gemini 3.1 Flash Image — launched on 26 Feb 2026, combining Nano Banana Pro quality with Flash speed; it became the default image model across the Gemini app, AI Mode, Lens, Ads and Flow and debuted at #1 in the Artificial Analysis text-to-image arena. GA as `gemini-3.1-flash-image` followed on 28 May.

- Preview 2026-02-26 as gemini-3.1-flash-image-preview; GA gemini-3.1-flash-image on 2026-05-28
- Default image engine in Gemini app, Search AI Mode, Google Lens, Google Ads and Flow
- Ranked #1 in Artificial Analysis Text-to-Image arena shortly after launch (per press)

#### What happened
Google released a faster, more realistic successor to its viral Nano Banana image model and made it the default image generator across its products.

#### Why it matters
Image generation/editing became a major driver of Gemini adoption (Google later reported 150M+ images generated daily in the Gemini app).

Sources: [Google: Nano Banana 2](https://blog.google/innovation-and-ai/technology/ai/nano-banana-2/); [TechCrunch: Google launches Nano Banana 2](https://techcrunch.com/2026/02/26/google-launches-nano-banana-2-model-with-faster-image-generation/); [Workspace Updates: Nano Banana 2 in the Gemini app](https://workspaceupdates.googleblog.com/2026/02/introducing-nano-banana-2-in-gemini-app.html); [Gemini API release notes](https://ai.google.dev/gemini-api/docs/changelog)

### 2026-03-17: Midjourney V8 alpha: rebuilt GPU-native model, ~5x faster, native 2K

**; Midjourney; Media generation; confidence medium; https://postcutoff.com/e/2026-03-17-midjourney-v8-alpha/

Midjourney released V8 as an alpha on 2026-03-17 — its first model on a completely new GPU/PyTorch codebase — with ~4-5x faster generation, native 2K 'HD' images and better text rendering; V8.1 (2026-04-14) became the default from June 10.

- V8.0 alpha launched 2026-03-17 on the Midjourney alpha site
- V8.1 released 2026-04-14; default version from 2026-06-10 to 2026-07-23 per Midjourney docs
- Standard jobs render about 4-5x faster than earlier versions; native 2K images without upscaling
- First Midjourney model on a new GPU-native codebase (moved off TPUs)

#### What happened
Midjourney's long-awaited V8 shipped first as an alpha, then V8.1, which restored a V7-like aesthetic with more stable moodboards and style references.

#### Why it matters
Midjourney remains the leading independent image generator; the platform rewrite lets it iterate faster against Google, OpenAI and Chinese rivals.

Sources: [Midjourney docs: Version](https://docs.midjourney.com/hc/en-us/articles/32199405667853-Version); [Midjourney updates: V8.1 Alpha](https://updates.midjourney.com/v8-1-alpha/)

### 2026-03-24: OpenAI shuts down Sora: app and web closed April 26, API on Sept 24, 2026; Disney's $1B deal falls through

***; OpenAI, Disney; Media generation; confidence high; https://postcutoff.com/e/2026-03-24-openai-shuts-down-sora/

On March 24, 2026 OpenAI announced it was "saying goodbye" to Sora, its AI video app launched six months earlier with Sora 2. The web and app experiences closed on April 26, 2026 and the Sora API on Sept 24, 2026, with no direct replacement. Press reports cite heavy compute costs (about $1M a day), falling usage and a refocus ahead of an IPO. Disney's planned $1B investment and character-licensing deal did not go ahead.

- Announcement on X, March 24, 2026: 'We're saying goodbye to Sora… We'll share more soon, including timelines for the app and API and details on preserving your work' (NBC News, TechCrunch)
- Timeline (OpenAI Help Center): Sora web and app discontinued April 26, 2026; Sora API (sora-2, sora-2-pro) discontinued Sept 24, 2026; exports via sora.chatgpt.com/sunset
- Usage (TechCrunch, citing app-intelligence data): downloads peaked at 3.3M in Nov 2025 and fell to 1.1M by Feb 2026; ~$2.1M lifetime in-app revenue
- Cost: Sora was reported to be losing about $1M a day in compute (TechCrunch, Apr 17)
- Disney's three-year partnership and planned $1B equity investment (announced Dec 2025, giving Sora access to 200+ Disney, Marvel, Pixar and Star Wars characters) will not proceed; Disney said it 'respect[s] OpenAI's decision to exit the video generation business' (NBC News)
- Reasons given in reporting: compute demands, a strategic decision that OpenAI cannot pursue 'everything at once', competition with Anthropic in enterprise, and IPO preparation; OpenAI's notice gave no explicit reason
- Follow-on: on April 17 Sora lead Bill Peebles left OpenAI as it cut 'side quests' (separate entry)

#### What happened
OpenAI closed its consumer video product half a year after Sora 2 and its social app topped the App Store. The shutdown came in two
stages: the consumer app on April 26 and the developer API on Sept 24, 2026. Resellers such as ElevenLabs removed Sora models at the
same time (see the sora-2 model page). TechCrunch reported at announcement time that the Sora 2 model remained available inside
ChatGPT's paid tiers. We have not confirmed whether that is still true after Sept 24.

#### Update, October 2026: talks to license Sora to a film startup
The Information reported on Oct 12, 2026, citing one source, that OpenAI is in talks with a startup that makes AI video tools for
filmmakers and wants to use Sora to sell software and services. The startup was co-founded by Jeffrey Katzenberg, former Sora
research lead Bill Peebles and former Dropbox CFO Sujay Jaswa; its name is not known. It is unclear whether it would receive Sora's
weights, but the source said it would get access deep enough to customise the model substantially. (Paywalled; from PYMNTS's summary.)

#### Why it matters
It is the clearest example of a frontier lab abandoning a flagship consumer product because of compute costs. OpenAI chose to spend its
GPUs on reasoning models, agents and enterprise work. It left the frontier video market to Google (Veo), ByteDance, Kuaishou and
Runway, and it marked the start of OpenAI's 2026 drive to drop "side quests" ahead of an IPO.

Sources: [OpenAI Help Center: What to know about the Sora discontinuation](https://help.openai.com/en/articles/20001152-what-to-know-about-the-sora-discontinuation); [NBC News: OpenAI is shutting down its Sora video-creation app](https://www.nbcnews.com/tech/tech-news/openai-shuttering-sora-video-generating-service-rcna264989); [TechCrunch: OpenAI's Sora was the creepiest app on your phone, now it's shutting down](https://techcrunch.com/2026/03/24/openais-sora-was-the-creepiest-app-on-your-phone-now-its-shutting-down/); [The Decoder: OpenAI sets two-stage Sora shutdown, app closing April 2026 and API in September](https://the-decoder.com/openai-sets-two-stage-sora-shutdown-with-app-closing-april-2026-and-api-following-in-september/); [The Hill: OpenAI shuts down Sora video app amid deepfake concerns](https://thehill.com/policy/technology/5799879-openai-sora-video-app-shutdown/); [OpenAI API deprecations](https://developers.openai.com/api/docs/deprecations); [PYMNTS: OpenAI's Video Model Sora Could Get a Second Act (Oct 12, 2026)](https://www.pymnts.com/news/artificial-intelligence/2026/openais-video-model-sora-could-get-a-second-act/); [The Information: OpenAI's Plan To Revive Sora (Oct 12, 2026)](https://www.theinformation.com/newsletters/ai-agenda/openais-plan-revive-sora)

### 2026-03-26: Suno v5.5 lets users sing with their own cloned voice and fine-tune personal models

**; Suno; Media generation; confidence high; https://postcutoff.com/e/2026-03-26-suno-v5-5-voices-custom-models/

Suno released v5.5, its last pre-licensing flagship, with three personalization features: Voices (verified cloning of the user's own singing voice), Custom Models (fine-tuning a private v5.5 on at least 6 of the user's own tracks) and My Taste (learned style preferences). It moved consumer AI music from "generic song" toward "your voice, your sound".

- Announced 2026-03-26 on Suno's blog (MBW dated the release Friday 2026-03-27)
- Voices: record/upload your own singing; a verification step has the user speak a random phrase to prove it is their voice; voices private by default; Pro/Premier only
- Custom Models: upload at least 6 tracks from your own catalog to tune v5.5 to your style; up to 3 custom models per user; Pro/Premier only
- My Taste: learns preferred genres, moods and references and applies them via the Magic Wand; all users
- v5.5 was retired on 2026-09-09 when Suno replaced its lineup with the licensed-data v6 family; Voices and Custom Models carried over

#### What happened
Suno shipped v5.5, billed as its "most expressive" and "most personal" model, with richer arrangements and sharper vocals than v5. The headline was personalization: paying users could capture their own singing voice (with an anti-impersonation verification step) and have Suno sing generated songs in it, and could fine-tune a private copy of v5.5 on their own catalog. Suno framed it as "The best music starts with a human."

#### Why it matters
It brought consumer-grade voice cloning and per-user fine-tuning into the most popular AI music app, raising both creative possibilities and impersonation/consent questions, six months before Suno retired all of its unlicensed-data models in favor of v6.

Sources: [Suno blog: v5.5 - More Expressive. More You.](https://about.suno.com/blog/v5-5); [Suno release notes: Introducing v5.5 - Voices, Custom Models, and My Taste](https://suno.com/release-notes/introducing-v5-5-voices-custom-models-and-my-taste); [Music Business Worldwide: Suno launches v5.5 AI model with voice cloning tool](https://www.musicbusinessworldwide.com/suno-launches-v5-5-ai-model-with-voice-capture-and-personalization-features/)

### 2026-05-19: Google unveils Gemini Omni, an any-to-any model that generates and conversationally edits video

****; Google DeepMind, Google; Media generation; confidence high; https://postcutoff.com/e/2026-05-19-gemini-omni/

Gemini Omni, announced at I/O on 19 May 2026, is Google's first "any-to-any" model family: Gemini Omni Flash takes text, images, audio and video in one prompt and outputs physics-aware video that can be edited turn-by-turn in plain language, with SynthID watermarks. It rolled out to paid Gemini/Flow users and free on YouTube Shorts; API access came 30 June and Omni 1.1 Flash on 27 Aug.

- Announced 2026-05-19 at Google I/O; blog authored by Koray Kavukcuoglu
- Inputs: any mix of text, image, audio, video; first release (Omni Flash) outputs video only — image and audio output promised later
- Conversational editing keeps characters, lighting and continuity across turns; avatars with your own voice
- Rolled out to Google AI Plus/Pro/Ultra in Gemini app and Flow; free in YouTube Shorts Remix and YouTube Create (18+)
- SynthID watermark on every clip; speech-editing of real people restricted
- Developer API (gemini-omni-flash-preview) launched 2026-06-30; reported ~$0.10 per second of generated video

#### What happened
Instead of a standalone "Veo 4", Google introduced **Gemini Omni**, a generative model family that reasons across modalities rather than stitching separate models together. Gemini Omni Flash accepts a portrait, a location photo, a voice sample and a one-line brief in a single prompt and returns a single coherent shot; follow-up prompts edit the same scene. It shipped to consumers the same day and to Google Vids (Workspace) in July.

#### Why it matters
Omni folds Google's generative media stack (Veo, Nano Banana, Genie-style world knowledge) into the Gemini model line, and shifts video generation from one-shot prompting to iterative, conversational editing — a workflow closer to real production.

Sources: [Introducing Gemini Omni (Google blog)](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-omni/); [Gemini Omni Flash model card](https://deepmind.google/models/model-cards/gemini-omni-flash/); [9to5Google: Gemini Omni, the 'create anything' model](https://9to5google.com/2026/05/19/gemini-omni-create-anything-model-video/); [TechCrunch: Gemini Omni turns images, audio and text into video](https://techcrunch.com/2026/05/19/googles-gemini-omni-turns-images-audio-and-text-into-video-and-thats-just-the-start/); [Introducing Gemini Omni: Create Anything from Anything (video)](https://www.youtube.com/watch?v=KUyRq7szZsM); [Gemini Omni Flash now in Google Vids (Workspace blog)](https://workspace.google.com/blog/product-announcements/introducing-gemini-omni-flash-in-google-vids)

### 2026-05-28: ElevenLabs Dubbing v2: direct speech-to-speech dubbing in 90+ languages

***; ElevenLabs; Media generation; confidence high; https://postcutoff.com/e/2026-05-28-elevenlabs-dubbing-v2/

On 2026-05-28 ElevenLabs introduced Dubbing v2, which conditions directly on the original speech instead of an ASR-translate-TTS pipeline, so emotion and performance carry across 90+ languages. The API followed in August 2026 at $2.20/min.

- Speech-to-speech architecture 'conditioning directly on the original performance'; 90+ languages
- ElevenLabs claim: 'For the first time, the emotion and performance of the original speaker carries across every language'
- UI launch 2026-05-28 (ElevenCreative, ElevenProductions); API 2026-08-06 (blog) / 2026-08-10 (changelog)
- API price $2.20/min (Dubbing v1: $0.33/min watermarked, $0.50 unwatermarked); docs label it 'Dubbing v2 Alpha'

#### What happened
ElevenLabs replaced its cascaded dubbing pipeline with a direct audio-to-audio model. Model file: `data/models/elevenlabs-dubbing-v2.md`.

#### Why it matters
End-to-end speech-to-speech translation that keeps each speaker's performance makes AI dubbing viable for expressive film and creator content. The "first" is a company claim.

Sources: [ElevenLabs blog: Introducing Dubbing v2](https://elevenlabs.io/blog/introducing-dubbing-v2); [ElevenLabs blog: Dubbing v2 API](https://elevenlabs.io/blog/dubbing-api); [Docs: Dubbing](https://elevenlabs.io/docs/overview/capabilities/dubbing); [API pricing](https://elevenlabs.io/pricing/api)

### 2026-06-30: Gemini Omni Flash opens to developers via the Gemini API

**; Google; Media generation; confidence high; https://postcutoff.com/e/2026-06-30-gemini-omni-flash-api/

On 30 June 2026 Google released `gemini-omni-flash-preview` in the Gemini API and AI Studio (plus `gemini-3.1-flash-lite-image` GA), letting developers generate and conversationally edit video with Gemini Omni for roughly $0.10 per second of output. The preview was superseded by Gemini Omni 1.1 Flash on 27 Aug.

- Model ID: gemini-omni-flash-preview (deprecated 2026-09-30 in favour of gemini-omni-1.1-flash)
- Pricing per Gemini API docs: $17.50 per 1M video output tokens, 5,792 tokens per second of 720p video (~$0.10/s)
- Same-day GA of gemini-3.1-flash-lite-image

#### What happened
Six weeks after its consumer debut, Gemini Omni Flash became available to developers through the Gemini API and Google AI Studio as a preview model.

#### Why it matters
API access turned Omni from a consumer feature into a building block; third-party creative tools began integrating it (Adobe Firefly, Figma Weave and Runway integrated the later 1.1 version).

Sources: [Gemini API release notes (30 June 2026)](https://ai.google.dev/gemini-api/docs/changelog); [Gemini API pricing](https://ai.google.dev/gemini-api/docs/pricing); [Gemini Omni Flash model card](https://deepmind.google/models/model-cards/gemini-omni-flash/)

### 2026-07-23: Black Forest Labs unveils FLUX 3: one model for images, 20-second video with audio, and robot actions

****; Black Forest Labs; Media generation; confidence high; https://postcutoff.com/e/2026-07-23-black-forest-labs-flux-3/

Germany's Black Forest Labs announced FLUX 3 on 2026-07-23, a multimodal flow model jointly trained on images, video, audio and action prediction; it is BFL's first video model (clips up to 20 s with synced audio) and powers FLUX-mimic, a robot-manipulation model being tested by Audi. A 7B open-weights FLUX 3 Action followed on 2026-09-23.

- Single architecture jointly trained on images, video, audio and action prediction
- FLUX 3 Video: clips up to 20 seconds with synchronized audio; aspect ratios 9:16 to 21:9; up to 10 image references (secondary sources)
- FLUX-mimic (with mimic robotics): fine-tunes to a task with ~30 minutes of robot data vs 30+ hours previously
- Audi testing FLUX-mimic for soft-body manipulation in production and logistics
- Launch partners/testers: Adobe Photoshop, Canva, Picsart, Krea, Burda, Magnific; Nous Research's Hermes Agent
- Video and Action in early access at launch; open-weight and faster versions promised later in 2026
- FLUX 3 Action: 7B open-weights robot-control model published 2026-09-23 (DataNorth)

#### What happened
Black Forest Labs (maker of FLUX image models) moved beyond still images with FLUX 3. The same backbone generates images, video with native audio, and robot action sequences.
Its robotics application, FLUX-mimic, built with Swiss startup mimic robotics, is claimed to cut the robot data needed for a new manipulation task from 30+ hours to ~30 minutes; Audi is deploying it in pilots.
FLUX 3 Video and Action launched in gated early access; on 2026-09-23 BFL published FLUX 3 Action as a 7B open-weights model.

#### Why it matters
FLUX 3 is a concrete instance of the "world model → robot policy" convergence: a generative video model doubling as a robot foundation model. It also makes BFL, a European lab, a full-stack video competitor.

Sources: [GlobeNewswire: Black Forest Labs unveils FLUX 3](https://www.globenewswire.com/news-release/2026/07/23/3332364/0/en/black-forest-labs-unveils-flux-3-a-new-multimodal-frontier-model-for-visual-intelligence.html); [BFL blog: FLUX 3 Video, Part 1: Generation](https://bfl.ai/blog/flux-3-video); [VentureBeat: FLUX 3 generates images and 20-second video with audio](https://venturebeat.com/technology/black-forest-labs-launches-flux-3-capable-of-generating-images-and-20-second-video-with-audio-but-in-limited-release-to-start); [MarkTechPost: FLUX 3 multimodal flow model](https://www.marktechpost.com/2026/07/26/black-forest-labs-releases-flux-3-a-multimodal-flow-model-for-image-video-audio-and-robot-action-prediction/); [DataNorth: FLUX 3 Action 7B robotics model](https://datanorth.ai/news/black-forest-labs-releases-flux-3-action)

### 2026-07-29: Google launches Lyria 3.5 music model in Flow Music; Gemini API GA follows

***; Google DeepMind, Google; Media generation; confidence high; https://postcutoff.com/e/2026-07-29-google-lyria-3-5/

Google DeepMind released Lyria 3.5, its third Lyria model in about five months, first in Google Flow Music, with better melodies, lyrics, more natural vocals and tempo/duration control; it became generally available in the Gemini API as lyria-3.5 on 2026-09-03 at $0.08 per full song.

- Launched 2026-07-29 in Google Flow Music (the former ProducerAI)
- Improvements: musicality, lyric quality and prompt adherence, vocal expressiveness and pronunciation, tempo and duration control
- Gemini API id lyria-3.5 (Stable, Interactions API), GA 2026-09-03; $0.08 per song, no free tier
- 44.1 kHz stereo MP3/WAV, text + image input, SynthID watermark
- Lyria 3 Clip/Pro previews now labelled legacy on the Gemini API pricing page

#### What happened
Lyria 3.5 replaced Lyria 3 Pro behind Google Flow Music's song generation on launch day, at no extra cost to Flow Music users. About five weeks later it reached general availability for developers in the Gemini API's Interactions API. As of 2026-09-29 it was not yet listed on Vertex AI (Gemini Enterprise Agent Platform), which still offers Lyria 3 previews and Lyria 2.

#### Why it matters
Google now ships a GA, watermarked, pay-per-song music model to developers, something Suno (web app only, API only "being explored") and Udio (no public API) do not offer.

Sources: [Google: Lyria 3.5 in Google Flow Music](https://blog.google/innovation-and-ai/models-and-research/google-labs/lyria-3-5/); [Lyria 3.5 model card](https://deepmind.google/models/model-cards/lyria-3-5/); [Gemini API music generation docs](https://ai.google.dev/gemini-api/docs/music-generation); [Gemini API pricing](https://ai.google.dev/gemini-api/docs/pricing); [Tech Times on Lyria 3.5](https://www.techtimes.com/articles/322113/20260729/googles-lyria-35-sharpens-vocals-lyrics-while-rivals-fight-court.htm)

### 2026-07-31: ByteDance releases Seedance 2.5: 30-second one-take audio-video clips with up to 50 references; a 4K film-production version follows on Oct 23

***; ByteDance; Media generation; confidence high; https://postcutoff.com/e/2026-07-31-seedance-2-5/

On July 31, 2026 ByteDance Seed released Seedance 2.5, a joint audio-video generation model that makes a 30-second clip with synced sound in one pass (up from 15 s) and takes up to 30 images, 10 video clips and 10 audio clips as references. It became one of the most used video models in AI-made media; on Oct 23 Volcano Engine opened a 4K, 10-bit 'Seedance 2.5 4K' version to selected film and TV customers.

- Announced July 31, 2026 by ByteDance Seed; rolled out in Jimeng AI and Doubao Pro, with API access via BytePlus ModelArk 'coming soon' at launch (ModelArk id dreamina-seedance-2-5-260628, see the model file)
- Up to 30 s per generation in one take (previously 15 s), extendable over several rounds while keeping characters, setting and narrative consistent
- Unified audio-video joint generation: sound and picture made together and kept in sync
- References in a single pass: up to 30 images, 10 video clips and 10 audio clips; also clay-render, motion and creative references
- Editing: timestamp-level targeted audio/video edits, green screen, camera-perspective and reference-based editing
- Seedance 2.5 4K (Oct 23, 2026, Volcano Engine): 4K output with 10-bit colour, further training for facial expressions and fine motion, a 480p draft mode for cheap storyboard previews; trials for selected enterprise customers only, some Chinese and overseas film and post-production firms testing (TechNode citing IT Home); no price given
- Artificial Analysis text-to-video leaderboard (read 2026-10-25): #5, Elo 1158
- Widely used in the AI-made videos in this dataset (e.g. the Claude Pop genre)

#### What happened
ByteDance's Seed team released Seedance 2.5 on July 31, 2026. It doubles the single-pass length to 30 seconds and generates sound
and picture together, so dialogue and effects stay in sync. A single prompt can carry up to 50 reference files (images, clips and
audio), and edits can target a time range in the clip. It launched in ByteDance's consumer apps (Jimeng, Doubao) with the API
following on BytePlus ModelArk.

On Oct 23, 2026 ByteDance's cloud arm Volcano Engine opened a 4K version for film and television work, with 10-bit colour and a
low-cost 480p draft mode for testing shots. Only selected enterprise customers can try it so far.

#### Why it matters
Seedance 2.5 is one of the two or three video models most creators used in the second half of 2026, and the 4K version marks a push
from social clips toward professional production. Third-party pages had claimed native 4K for the July model, but BytePlus pricing
listed 4K as unsupported; the October version is the first official 4K release.

Sources: [ByteDance Seed: Introducing Seedance 2.5 (July 31)](https://seed.bytedance.com/en/blog/one-take-creation-flexible-referencing-introducing-seedance-2-5); [TechNode: ByteDance launches Seedance 2.5 video generation model (July 31)](https://technode.com/2026/07/31/bytedance-launches-seedance-2-5-video-generation-model/); [TechNode: ByteDance launches Seedance 2.5 4K video model for film production (Oct 23)](https://technode.com/2026/10/23/bytedance-launches-seedance-2-5-4k-video-model-for-film-production/)

### 2026-08-11: Lightricks releases LTX-2.5, a 22B open-weights audio-video 'world model' with multishot generation and up to 4K output

***; Lightricks; Media generation; confidence high; https://postcutoff.com/e/2026-08-11-lightricks-ltx-2-5-open-weights/

On Aug 11, 2026 Lightricks released LTX-2.5, a 22B-parameter open-weights model that generates synchronized video and audio from text, images or video, keeps characters and voices consistent across cuts (multishot), and ships a checkpoint for robotics and physical AI. Free for organisations under $10M revenue, it became the most-downloaded generative model on Hugging Face (~1.6M downloads a month by October).

- 22B DiT in a full (trainable) and a distilled 8-step variant; Gemma 4 12B text encoder; new diffusion video decoder (Hugging Face model card)
- Speed: a 10-second 720p image-to-video clip in 6.8 s on two Nvidia GB200 superchips (Lightricks via VentureBeat)
- Up to 4K (3840×2176); clips up to 20 s at 24/25 fps through the API; native audio with an integrated vocoder
- LTX-2.x Community License: free for commercial use under $10M annual revenue, paid licence above; use for military/weapons development is banned
- Managed API: Fast tier $0.09/s, 4K $0.30/s (VentureBeat)
- Lightricks cites 33M+ cumulative downloads for the LTX family; LTX-2.5 had ~1.58M monthly downloads and topped Hugging Face weekly trending on Oct 1, 2026
- CEO Zeev Farbman: 'With video models, world models, there are so many different use cases that require people to get access to the weights'

#### What happened
Lightricks, the Israeli maker of Facetune and LTX Studio, released its third major open video model generation. LTX-2.5 runs locally, has day-one ComfyUI support, and is pitched both for creative video and as a world model for real-time applications and robotics.

#### Why it matters
Most frontier video models (Veo, Sora, Kling) are closed. LTX-2.5 is among the strongest open audio-video models anyone can fine-tune, and its download numbers make it a default base for the open video community.

Sources: [Hugging Face: Lightricks/LTX-2.5 model card](https://huggingface.co/Lightricks/LTX-2.5); [GitHub: Lightricks/LTX-2](https://github.com/Lightricks/LTX-2); [VentureBeat: LTX-2.5 can generate a 10-second AI video from an image in 6.8 seconds](https://venturebeat.com/technology/ltx-2-5-can-generate-a-10-second-ai-video-from-an-image-in-just-6-8-seconds-on-nvidia-superchips-and-its-open-weights); [ComfyUI Wiki: LTX-2.5 open-weights release](https://comfyui-wiki.com/en/news/2026-08-11-ltx-2-5-open-weights-release); [Runpod: running LTX-2.5](https://www.runpod.io/blog/ltx-2-5-the-open-weights-world-model-built-for-speed-and-how-to-run-it-on-runpod)

### 2026-08-24: Alibaba launches Wan 3.0: closed-weight video model with 30-second 1080p clips, native audio and document or web-page inputs

***; Alibaba; Media generation; confidence high; https://postcutoff.com/e/2026-08-24-wan-3-0/

On Aug 24, 2026 Alibaba Cloud made Wan 3.0 generally available on Model Studio: up to 30-second clips at up to 1080p in one pass, native sound and multilingual voices, and inputs from text, images, audio, video, PDFs, slides and web pages. Unlike Wan 2.2, the weights are closed. By late October it ranked #1 on Artificial Analysis' text-to-video leaderboard.

- GA on Alibaba Cloud Model Studio Aug 24, 2026 after a public beta from early August (application-based); API callable without applying
- Up to 30 s per generation (Wan 2.7: 15 s), up to 1080p; consistent characters, props, layout and motion graphics across the clip
- Native audio with synced facial micro-expressions and multilingual voice; inputs: text, images, audio, video, documents (PDF, PowerPoint) and web pages; up to 10 images, 5 videos and 5 audio clips as references
- Price per second: CNY 0.3 / 0.6 / 1.2 for 480p / 720p / 1080p (TNW: $0.05 / $0.10 / $0.20); failed generations are free
- Closed weights: Wan 2.2 (July 2025) was the last open-sourced Wan flagship (TNW)
- Context (TNW): launched days after Alibaba's ~$10.2B Hong Kong share placement earmarked for AI
- Artificial Analysis text-to-video leaderboard (read 2026-10-25): #1, Elo 1175

#### What happened
Alibaba's Tongyi Lab moved Wan 3.0 from beta to general availability on Aug 24, 2026. The model doubles clip length to 30 seconds,
adds sound and voices generated with the picture, and accepts unusual inputs such as slide decks and web pages, which it turns into
video. It is sold per second through Model Studio's API.

#### Why it matters
Wan had been the leading open-weight video family; Wan 3.0 keeps its weights closed, part of a wider 2026 shift by Chinese labs
toward closed flagships. It later topped the main independent text-to-video leaderboard.

Sources: [Alibaba Cloud: video generation model Wan 3.0 officially launched](https://www.alibabacloud.com/en/news/product/video-generation-model-wan30-officially-launched-nsb); [The Next Web: Alibaba launches Wan 3.0 video model after share sale (Aug 24)](https://thenextweb.com/news/alibaba-wan3-video-model-after-share-sale)

### 2026-08-27: Gemini Omni 1.1 Flash adds scene extension, frame interpolation and 4K upscaling

***; Google DeepMind, Google; Media generation; confidence high; https://postcutoff.com/e/2026-08-27-gemini-omni-1-1-flash/

Google made Gemini Omni 1.1 Flash (`gemini-omni-1.1-flash`) generally available on 27 Aug 2026, adding scene extension up to 40 s, first/last-frame interpolation, 1080p and 4K output, and cheap 360p drafts; Adobe Firefly, Figma Weave and Runway integrated it.

- GA 2026-08-27; model ID gemini-omni-1.1-flash; gemini-omni-flash-preview deprecated 2026-09-30
- Scene extension up to 40 seconds total, using up to 10 s of prior context (previously 1 s)
- First-and-last-frame interpolation; video references up to 3 s
- Output 1080p and 4K (upscaling); 360p drafts up to 60% faster at one third the cost of 720p
- Available in AI Studio, Gemini Enterprise Agent Platform, Google Flow (AI Plus/Pro/Ultra) and the Gemini app
- Integrated by Adobe Firefly, Figma Weave and Runway

#### What happened
Gemini Omni 1.1 Flash reached general availability with production-oriented controls: extending scenes with continuity, specifying first and last frames, 4K upscaling and fast low-resolution previews for iteration.

#### Why it matters
These are the controls professional video workflows need (continuity, shot planning, resolution), and adoption by Adobe, Figma and Runway puts Google's model inside mainstream creative tools.

Sources: [Build with Gemini Omni 1.1 Flash (Google blog)](https://blog.google/innovation-and-ai/technology/developers-tools/build-with-gemini-omni-1-1-flash/); [Gemini API release notes (27 Aug 2026)](https://ai.google.dev/gemini-api/docs/changelog); [Google AI announcements from August 2026](https://blog.google/innovation-and-ai/technology/google-ai-updates-august-2026/); [Google DeepMind: Gemini Omni 1.1 Flash lets you build with more control](https://deepmind.google/blog/gemini-omni-1-1-flash-lets-you-build-with-more-control/); [Google blog: Anyone can make stunning HD videos with Gemini Omni in Google Vids (Sept 23)](https://blog.google/products-and-platforms/products/workspace/gemini-omni-in-google-vids/)

### 2026-09-04: Microsoft releases MAI-Image-2.6 and MAI-Image-2.6-Flash in Foundry; MAI-Image-2.6 debuts at No. 2 on Arena

**; Microsoft; Media generation; confidence high; https://postcutoff.com/e/2026-09-04-microsoft-mai-image-2-6/

On Sept 4, 2026 Microsoft AI released its in-house image models MAI-Image-2.6 and the faster MAI-Image-2.6-Flash in public preview on Microsoft Foundry. MAI-Image-2.6 ranked No. 2 on the Arena text-to-image leaderboard, 79 Elo above MAI-Image-2.5 (91 Elo in text rendering), and adds precise image editing and Bing web grounding.

- Model card dated Sept 4, 2026; public preview on Microsoft Foundry
- Arena text-to-image: No. 2 at launch, +79 Elo overall and +91 Elo on text rendering vs MAI-Image-2.5 (Microsoft)
- Diffusion-based; text-to-image plus controllable image-to-image editing; web grounding via Bing Search
- MAI-Image-2.6-Flash: 2.8x faster than GPT-Image-2-Medium with '72% greater efficiency' (Microsoft claim)

#### What happened
Microsoft AI (Mustafa Suleyman's group) shipped the next version of its own image model family through Foundry, with a premium and a Flash tier.

#### Why it matters
It is part of Microsoft's push to build first-party models alongside its OpenAI partnership. A top-two Arena placing puts its in-house image model among the leaders.

Unverified: some secondary sources date a first MAI-Image-2.6 announcement to Aug 10, 2026; this entry uses the Sept 4 model-card and Foundry release date.

Sources: [Microsoft AI: Pushing the quality-cost frontier with MAI-Image-2.6](https://microsoft.ai/news/pushing-the-quality-cost-frontier-with-mai-image-2-6/); [Microsoft AI: MAI-Image-2.6 launches at No. 2 on Arena](https://microsoft.ai/news/mai-image-2-6-launches-at-no-2-on-arena-ahead-of-google-meta-and-xai/); [MAI-Image-2.6 model card (PDF)](https://microsoft.ai/pdf/MAI-Image-2.6-Model-Card.pdf); [Microsoft Foundry blog: MAI-Image-2.6 and MAI-Image-2.6-Flash](https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/mai-image-2-6-and-mai-image-2-6-flash-quality-and-speed-at-production-scale/4550970)

### 2026-09-08: OpenAI launches ChatGPT Images 2.5 with Sketch, plus GPT-Image-2.5 Flare and Sunburst in the API

***; OpenAI; Media generation; confidence high; https://postcutoff.com/e/2026-09-08-chatgpt-images-2-5/

On Sept 8, 2026 OpenAI released ChatGPT Images 2.5, a new image model that it says gives sharper detail, keeps subjects from reference photos more faithfully, edits more precisely over many turns, and cuts latency by up to 50% versus Images 2.0. ChatGPT gained Sketch (draw a rough layout as the reference), templates, comments on images and shareable prompts. Developers got two API models: gpt-image-2.5-flare, the fast default, and gpt-image-2.5-sunburst, a slower, more precise premium model.

- OpenAI: people create more than 3 billion images per week across ChatGPT Images and GPT-Image models in the API
- Up to 50% lower generation latency than Images 2.0; better reference-photo fidelity, targeted edits that leave the rest of the image unchanged, multi-turn edit consistency, more accurate real-world content, complex layouts incl. transparent backgrounds
- ChatGPT features: Sketch ('@Sketch' drawing surface as a visual guide), templates (e.g. Poster, Merch), comments placed directly on images, sharing the prompt with an image
- Available to all ChatGPT, ChatGPT Work and Codex users on all tiers (desktop, mobile, web)
- API: gpt-image-2.5-flare (default; higher quality than gpt-image-2 at 50% lower latency) and gpt-image-2.5-sunburst (premium, tighter control, longer generation times); snapshot gpt-image-2.5-flare-2026-09-08
- API token prices listed identically for both models and gpt-image-2 (per 1M tokens): text input $5, cached $1.25; image input $8, cached $2; image output $30 (OpenAI pricing page, per our model registry)
- Safety: prompt and image checks, C2PA metadata and invisible watermarking; a system card on OpenAI's Deployment Safety Hub

#### What happened
OpenAI shipped **Images 2.5**, the successor to Images 2.0, in ChatGPT, ChatGPT Work and Codex for all users. It focuses on faithful
edits: keeping recognizable subjects from reference photos, changing only what was asked, and staying consistent across long editing
sessions. New tools include **Sketch**, which lets users draw a rough layout that ChatGPT turns into a finished image, format templates,
on-image comments and shareable prompts. In the API, OpenAI released two models: **GPT-Image-2.5 Flare** for most uses and
**GPT-Image-2.5 Sunburst** for premium work that needs tighter control.

#### Why it matters
Image generation in chat is now mostly about editing and control rather than one-off pictures, and OpenAI's own figure of more than
3 billion images a week shows how heavily it is used. Splitting the API into a fast and a precise model matches how Google and others
now tier their image models.

Sources: [OpenAI: Introducing ChatGPT Images 2.5](https://openai.com/index/introducing-chatgpt-images-2-5/); [OpenAI Deployment Safety Hub: ChatGPT Images 2.5 system card](https://deploymentsafety.openai.com/chatgpt-images-2-5); [OpenAI Developer Community: Introducing GPT Images 2.5 in the API and ChatGPT](https://community.openai.com/t/introducing-gpt-images-2-5-in-the-api-and-chatgpt/1395897); [OpenAI API docs: gpt-image-2.5-flare](https://developers.openai.com/api/docs/models/gpt-image-2.5-flare); [OpenAI API pricing (image generation)](https://developers.openai.com/api/docs/pricing#image-generation); [Axios: Hands-on with ChatGPT's new image editor](https://www.axios.com/2026/09/08/exclusive-hands-on-with-chatgpts-new-image-editor); [Unite.AI: OpenAI releases ChatGPT Images 2.5 with Sketch and two new API models](https://www.unite.ai/openai-releases-chatgpt-images-2-5-with-sketch-and-two-new-api-models/); [DataCamp: ChatGPT Images 2.5 features, Sketch and API models](https://www.datacamp.com/blog/chatgpt-images-2-5)

### 2026-09-09: Suno launches v6, its first music models trained on licensed music

***; Suno, Warner Music Group, BMG, Believe; Media generation; confidence high; https://postcutoff.com/e/2026-09-09-suno-v6-licensed-music-model/

On 2026-09-09 Suno launched the v6 family (v6, v6-wild, v6-mini), trained from scratch on music licensed from Warner Music Group, BMG and Believe with revenue sharing, and retired all older models; Sony Music and Universal sued again on 2026-09-18, alleging v6 was trained on outputs of the old unlicensed models.

- Three models: v6 (flagship, paid), v6-wild (more varied, paid), v6-mini (fast, free tier)
- Licensing partners: Warner (deal Nov 25 2025, settling its suit), BMG (Aug 12 2026), Believe/TuneCore (Sept 8 2026)
- All earlier models (v4 through v5.5) retired on launch day
- New features: section editing by prompt/lyrics, text/image/video references, stem separation, opt-in artist remixing
- Suno has raised >$819M (PitchBook via TechCrunch)
- Sony Music and UMG filed new suit in Massachusetts federal court on 2026-09-18

#### What happened
Suno, the largest AI music generator, replaced its entire lineup with a licensed-data model family. Partners receive a share of revenue from launch day and distribute it to rights holders.
Suno says v6 was not trained on the data used for earlier versions (which had included YouTube audio). The remaining majors, Sony and Universal, plus artist Jason Isbell, continue to litigate.

#### Why it matters
v6 is the clearest test yet of a licensed-training business model for generative media; the new Sony/UMG suit tests whether "clean-room" retraining on licensed data (but with learnings from older models) is enough.

Sources: [TechCrunch: Suno replaces its AI models with one trained on licensed music](https://techcrunch.com/2026/09/09/suno-replaces-its-ai-models-with-a-new-one-trained-on-licensed-music-as-copyright-suits-pile-up/); [Digital Music News: Suno launches v6](https://www.digitalmusicnews.com/2026/09/09/suno-v6-launch/); [Music Ally: Suno v6 — what you need to know](https://musically.com/2026/09/09/suno-launches-its-v6-ai-music-models-heres-what-you-need-to-know/); [MBW: Suno inks global licensing deal with BMG (Aug 2026)](https://www.musicbusinessworldwide.com/suno-inks-global-licensing-deal-with-bmg); [MBW: Suno inks global licensing deal with Believe (Sept 2026)](https://www.musicbusinessworldwide.com/suno-inks-global-licensing-deal-with-believe/)

### 2026-09-11: ElevenLabs releases Music v2.5

***; ElevenLabs; Media generation; confidence high; https://postcutoff.com/e/2026-09-11-elevenlabs-music-v2-5/

ElevenLabs released Music v2.5 (music_v2_5) on 2026-09-11, its most advanced text-to-music model. It has richer melodies and more live-sounding instruments, was preferred over v2 in a blind test of 47,885 pairs, and is available in ElevenMusic, ElevenCreative and the API at $0.15/min.

- API model id music_v2_5; $0.15 per minute of generated music
- Blind test on 47,885 paired samples: v2.5 preferred in the majority; largest gains in R&B/soul, hip hop/trap, rock/metal, orchestral/cinematic
- New default for prompted and reference-audio generation in ElevenCreative
- Commercial use allowed; lossless downloads: Free 5/day, Pro 400/month; tracks based on other artists' songs cannot be downloaded
- API support with 6,132-character composition chunks rolled out 2026-09-14

#### What happened
ElevenLabs shipped Music v2.5 as the new default music model across ElevenMusic (elevenmusic.io), ElevenCreative and the API. Model file: `data/models/elevenlabs-music-v2-5.md`.

#### Why it matters
It is a licensed-by-design competitor to Suno and Udio. The download protections were built with labels and publishers.

Sources: [ElevenLabs blog: Music v2.5](https://elevenlabs.io/blog/music-v2-5-model); [Docs: Models](https://elevenlabs.io/docs/models); [Changelog 2026-09-14](https://elevenlabs.io/docs/changelog); [YouTube (ElevenLabs): Introducing Music v2.5](https://www.youtube.com/watch?v=zXlVQ8rMJM0)

### 2026-09-20: Alibaba releases Qwen-Image-2.1, a 7B open-weights image generation and editing model with native transparency

***; Alibaba, Qwen; Media generation; confidence high; https://postcutoff.com/e/2026-09-20-qwen-image-2-1/

On Sept 20, 2026 Alibaba's Qwen team released Qwen-Image-2.1 with open weights: a unified text-to-image and image-editing model whose visual generator has just 7B parameters (32 single-stream DiT layers). It natively generates and edits transparent RGBA layers, takes up to 10 reference images and, per Qwen, "outperforms most closed-source models". It shipped with day-0 support in Diffusers, ComfyUI, vLLM-Omni and SGLang.

- Oct 9, 2026: Qwen released Qwen-Image-2.1-Turbo, an 8-step accelerated checkpoint (CFG=1, prefix KV caching) for generation and editing, same 7B architecture and Qwen Research license (Hugging Face model card)
- 7B parameters in the visual generation component (32 single-stream DiT layers); mixed-granularity attention and prefix KV-cache reuse
- Native RGBA: generate transparent images, edit transparent layers, extract subjects from photos
- Editing with up to 10 reference images; local edits via circles, painted annotations or masks; identity preservation for people and products
- License: Qwen Research License (not Apache); weights on Hugging Face and ModelScope
- Day-0 support: Diffusers (QwenImage21Pipeline), ComfyUI, vLLM-Omni, SGLang, LightX2V
- Qwen's launch post reached ~2.3M views on X; commentators claimed it beats Nano Banana 2 (a community claim, not an official benchmark)
- Released two months after Alibaba's closed Qwen-Image-3.0

#### What happened
Qwen returned to open weights for image generation after keeping Qwen-Image-3.0 closed, releasing a small model aimed at speed and at production
tasks such as compositing, product shots, infographics and virtual try-on.

#### Why it matters
A 7B open model with editing, multi-reference input and native transparency makes high-quality local image generation practical on a single GPU,
and it narrows the gap to the closed leaders.

Sources: [Hugging Face: Qwen-Image-2.1-Turbo model card (Oct 9, 2026)](https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo); [Qwen blog: Qwen-Image-2.1](https://qwen.ai/blog?id=qwen-image-2.1); [GitHub: QwenLM/Qwen-Image-2.1](https://github.com/QwenLM/Qwen-Image-2.1); [Hugging Face: Qwen/Qwen-Image-2.1](https://huggingface.co/Qwen/Qwen-Image-2.1); [Qwen on X: Meet Qwen-Image-2.1](https://x.com/Alibaba_Qwen/status/2101659302792679789)

### 2026-09-21: ElevenLabs Studio 4.0 turns ElevenCreative into an agentic AI video editor

**; ElevenLabs; Media generation; confidence high; https://postcutoff.com/e/2026-09-21-elevenlabs-studio-4/

On 2026-09-21 ElevenLabs released Studio 4.0 in ElevenCreative: an audio/video editor that generates video, images, voiceovers, music and sound effects on the timeline, with a "Studio Agent" co-editor that drafts a first cut from a text description. It extends ElevenLabs from voice into multi-model video production.

- Studio Agent: AI co-editor that 'drafts a first cut on the timeline - placing clips, generating voiceovers, and syncing sound effects' (web only)
- Generation of video, images, voice, music and SFX inside a project; redesigned timeline with frame-level zoom and clip snapping; captions as timeline clips; clip-level comments; rebuilt playback engine
- Available on every plan incl. Free (3 projects, watermarked video); paid plans from $6 Starter (per secondary coverage)
- Visual generation comes from third-party models that ElevenLabs hosts through its Image & Video API: Seedance 2.0/2.5, Veo 3.1, GPT Image 1-2.5, Nano Banana family and Seedream 5 per the docs; GPT Image 2.5 Flare/Sunburst added 2026-09-21. Sora 2 was removed on 2026-09-23 after OpenAI shut down the Sora API on 2026-09-24
- Same month: Eleven Music v2.5 (09-11), Reception AI receptionist (09-16), Eleven v4 TTS (09-28)

#### What happened
ElevenLabs rebuilt Studio, its long-form audio and video editor, around generation and an in-editor agent. A user describes a video, and Studio Agent places generated clips, voiceovers and sound effects on the timeline for manual refinement.

#### Why it matters
ElevenLabs had been a voice-model company. Studio 4.0 makes it a multi-model video production tool that pairs third-party video models with its own voice and music. That puts it in competition with CapCut, Descript and the video labs' own editors.

Sources: [ElevenLabs blog: Studio 4.0, the AI-native video editor in ElevenCreative](https://elevenlabs.io/blog/introducing-studio-4); [YouTube (ElevenLabs): Introducing Studio 4.0, the agentic video editor in ElevenCreative](https://www.youtube.com/watch?v=P-OZwbegYss); [ElevenLabs docs: Image & Video capabilities (model list)](https://elevenlabs.io/docs/overview/capabilities/image-video); [ElevenLabs changelog (2026-09-21 / 2026-09-23)](https://elevenlabs.io/docs/changelog)

### 2026-09-22: Tencent Hunyuan releases Hy Image 3.5 preview, an image model priced at ¥0.15 per 2K image

**; Tencent; Media generation; confidence medium; https://postcutoff.com/e/2026-09-22-tencent-hy-image-3-5-preview/

On Sept 22, 2026 Tencent Hunyuan released Hy Image 3.5 preview, a text-to-image and image-to-image model with up to five reference images, multi-turn editing and output up to 2K. Tencent priced it at ¥0.15 (international: $0.024) per 2K image on Tencent Cloud. In Tencent's internal designer tests it was about 30% better than Hy Image 3.0 and roughly on par with ByteDance's Seedream 5.0 Pro.

- Released Sept 22, 2026 (preview); text-to-image and image-to-image, up to 5 reference images, multi-round editing, various aspect ratios, up to 2K
- Tencent Cloud API: ¥0.15 per 2K image (international $0.024); billed per output image
- Tencent's internal evaluation by hundreds of professional designers: ~30% better than Hy Image 3.0, on par with Seedream 5.0 Pro, slightly better than Nano Banana Pro and Qwen-Image-3.0 Pro (company claim)
- Available in Yuanbao, WorkBuddy, ima and other Tencent apps, plus Tencent Cloud MPS and VOD

#### What happened
Tencent's Hunyuan team shipped a preview of its next image model, aimed at professional design work, at a low per-image API price.

#### Why it matters
Part of the 2026 price war among Chinese image models (Seedream, Qwen-Image, Hunyuan). All the quality comparisons are Tencent's own
internal tests, with no independent arena results yet.

Unverified: whether the weights will be open, as earlier HunyuanImage versions were; no Artificial Analysis or LMArena ranking checked.

Sources: [Tencent News: 腾讯混元发布最新生图模型 Hy Image3.5 preview](https://news.qq.com/rain/a/20260922A049W000); [Futu News: Tencent Hunyuan releases preview of Hunyuan Image 3.5](https://news.futunn.com/post/79588420/tencent-hunyuan-releases-preview-of-hunyuan-image-3-5-to); [NetEase: 腾讯混元发布混元图像3.5预览版](https://www.163.com/dy/article/L7E5ENOP05198UNI.html)

### 2026-09-28: Kuaishou's Kling unveils Kling 4.0: 30-second clips, 10 keyframes, ahead of possible HK listing

***; Kuaishou, Kling AI; Media generation; confidence high; https://postcutoff.com/e/2026-09-28-kling-4-0/

Kling AI, Kuaishou's video-generation spinoff, unveiled Kling 4.0 on 2026-09-28: it doubles maximum clip length to 30 seconds, accepts more than a dozen reference inputs (text, images, video) and up to 10 keyframes; a Lite version launched for annual subscribers with full rollout planned for October.

- Max clip length 30 s (up from 15 s in Kling 3.0)
- More than a dozen reference inputs across text, images and existing video; up to 10 keyframes
- Kling raised $2.8B in July 2026 at ~ $18B valuation; annualized revenue passed $500M by March 2026
- Preparing for a possible Hong Kong listing; Kuaishou retains majority stake
- Oct 16, 2026: Kling 4.0 went fully live on klingai.com with audio, MCP and CLI hooks (Pandaily headline; article body not readable for us). TechNode (Oct 16): up to 30 s per generation plus video extension, up to 15 reference assets (images, video, audio) and up to 10 keyframes; framed as a shift in AI video competition from visual quality to directorial control

#### What happened
Kling 4.0 arrived as Kling competes with ByteDance's Seedance (Seedance 2.5 also targets 30-second single-shot generation) and tools from Alibaba and MiniMax.
Kling is now run as an independent company that raised $2.8B in July and is weighing a Hong Kong IPO.

#### Why it matters
30-second coherent clips with keyframe control move AI video from short shots toward full scenes; Chinese companies (Kling, Seedance, Wan, Hailuo) now lead many video leaderboards, as TechCrunch noted in July.

Sources: [Bloomberg: Kuaishou's AI video spinoff unveils new model](https://www.bloomberg.com/news/articles/2026-09-28/kuaishou-s-ai-video-spinoff-unveils-new-model-in-bytedance-chase); [Briefs: Kling unveils 4.0 video model as Hong Kong listing nears](https://www.briefs.co/news/kuaishou-s-kling-unveils-4-0-video-model-as-hong-kong-listin/); [Kling AI blog](https://kling.ai/blog); [Pandaily: Kling 4.0 goes fully live on klingai.com with audio, MCP and CLI hooks](https://pandaily.com/kling-ai-kling-4-0-full-launch-klingai-com-audio-mcp-cli); [TechNode: Kling AI 4.0 suggests a shift from visual quality to directorial control](https://technode.com/2026/10/16/kling-ai-4-0-suggests-a-shift-in-ai-video-competition-from-visual-quality-to-directorial-control/)

### 2026-09-30: HeyGen releases HeyGen Video 1, a text/image/reference-to-video model with dialogue and sound, at $0.01 per second at launch

**; HeyGen; Media generation; confidence medium; https://postcutoff.com/e/2026-09-30-heygen-video-1/

HeyGen, known for AI avatar videos, released its own general video model, HeyGen Video 1, around Sept 30, 2026. It generates visuals, dialogue and sound effects in one pass from text, images or references, from 480p to 2K through the API, at a launch price of $0.01 per second. It ranked #3 on Artificial Analysis' text-to-video board in late October.

- Release date not printed by HeyGen; Sept 30, 2026 is OpenRouter's listing date and HeyGen's changelog says 'September 2026'
- Text-, image- and reference-to-video with dialogue and sound effects generated together; styles from live action to claymation and anime
- API: 480p to 2K, 16:9 / 9:16 / 1:1, seeds and webhooks; free web demo (two 5 s 768p clips a day)
- Launch API price $0.01 per second 'through October'
- HeyGen's page says it ranked first on Artificial Analysis' video leaderboard; read on 2026-10-25 it was #3, Elo 1160

#### What happened
HeyGen moved beyond avatar videos with a general video model that writes and voices its own dialogue. It is available in the HeyGen
app, as a free demo and through HeyGen's API and OpenRouter at a promotional price of one cent per second.

#### Why it matters
A cheap, top-five video model from an application company shows how quickly video generation has commoditised in 2026.

Sources: [HeyGen: Introducing HeyGen Video](https://www.heygen.com/blog/introducing-heygen-video); [OpenRouter: HeyGen Video 1](https://openrouter.ai/heygen/heygen-video-1)

### 2026-09-30: Utopai Studios launches its PAI production platform and Utopai X video model, #2 on Artificial Analysis' text-to-video board

**; Utopai Studios; Media generation; confidence medium; https://postcutoff.com/e/2026-09-30-utopai-x-pai/

On Sept 30, 2026 AI film studio Utopai Studios introduced PAI, a "production intelligence" platform that breaks scripts into scenes and shots and runs AI production assistants, with its own video model Utopai X inside. Utopai says Utopai X ranked #2 on Artificial Analysis' text-to-video-with-audio leaderboard (Elo 1150), the top US model. It has no public API.

- Announced Sept 30, 2026 (Utopai news page, Variety feature)
- PAI: scripts split into scenes and shots with linked characters, locations and visual references; agent 'Production Assistants' plan shots and produce iterations; Utopai X generates and compares takes that move into an editing timeline
- Claimed at launch: #2 on Artificial Analysis' Text-to-Video Leaderboard With Audio, Elo 1150, nine points behind Wan 3.0; still #2 (Elo 1167) when read on 2026-10-25, listed with 'No API available'
- Tested on Utopai's own productions, e.g. the animated feature 'The Most Serious Fart' (writer-director Mike Bender); Utopai seeks strategic partners
- Utopai calls itself the largest independent AI-native film and TV studio, valued at $1B; on Oct 24 The Information reported a $60M raise at a $1B valuation and 20+ projects in the works (paywalled; headline only)
- Quote, chief scientific officer Zijian He: 'Technology is most useful when it gives filmmakers more control over their ideas.'

#### What happened
Utopai Studios, which makes films and series with AI, released the platform it uses internally: a planning tool that turns a script
into a shot list and lets agents and its own video model generate takes for each shot. The model, Utopai X, placed second on the
leading blind-preference video leaderboard, which Utopai presented as the best result by a US company. It is not sold as an API.

#### Why it matters
A studio, not a lab, built a frontier-level video model for its own pipeline, a sign that video generation is moving into vertical
production tools.

Sources: [Utopai Studios: defining production intelligence for film and television](https://www.utopaistudios.com/en/news/utopai-studios-defining-production-intelligence-for-film-and-television); [Artificial Analysis text-to-video leaderboard](https://artificialanalysis.ai/video/leaderboard/text-to-video); [The Information: Utopai becomes a unicorn as video AI takes off (Oct 24, paywalled)](https://www.theinformation.com/articles/utopai-becomes-unicorn-video-ai-takes)

### 2026-10-01: Tavus unveils Griffin, a real-time face-to-face video model it says is the first to pass a 'video Turing test' (48% thought it was human)

***; Tavus; Media generation; confidence medium; https://postcutoff.com/e/2026-10-01-tavus-griffin-video-turing-test/

Tavus announced Griffin, a full-duplex video-to-video "Human Interaction Model" that watches and listens while it talks, can interrupt and be interrupted, and reacts to gestures and objects on camera. In Tavus's own blind study, 48% of 54 participants who had a one-minute live video call with it thought they had talked to a real person (its earlier systems: at most 2%). It is a research preview, not a general release. The launch post reached about 5.4M views on X.

- Blind study by Tavus: 54 participants, 1-minute live video calls; 48% judged Griffin human. Previous Tavus stack (Phoenix 4.5, Sparrow-2, Raven-1): max 2%
- NVIDIA VideoFDB (full-duplex video benchmark): Griffin-Lite generation score 3.83/5 vs 3.92 for the human reference and 2.80 for the next-best published system; perception 3.73
- Median response latency ~1.9 s vs ~0.9 s for the human reference (about 2.1×)
- Released as a research preview, not generally available
- Caveats: vendor-run study with a small sample and very short calls; 48% is just below the 50% a judge would score by guessing, so 'pass' depends on the definition used
- Reach and reactions on X: CEO Hassaan Raza's launch post ~3.35M views ('The first time I saw it being used, I had no idea I was watching our model'); Emad Mostaque ~570k ('Remote work is so cooked'); Kevin Roose ~400k, joking that Tavus built the 'trick old people into handing over their bank passwords machine'
- Backlash on Oct 2: Matt Walsh ~742k views ('a disaster for the human race… It will be used to scam old people'); @IterIntellectus ~906k ('they just killed like 20 entire countries' economies'); Sen. Bernie Sanders ~209k ('Does anyone in America think it is a good idea to race forward to build machines that replace us? PAUSE AI NOW.')
- Oct 3 follow-up by CEO Hassaan Raza: Griffin 'isn't publicly available yet', the announcement was 'a limited research preview'; Tavus is 'working with partners on safeguards and systems for disclosure' and invites discussion with officials on regulation; 'People will always know they're interacting with AI'
- Tavus says Griffin 'requires safety considerations before it can be released publicly' because it can be mistaken for a real person (Tavus on X)

#### What happened
Tavus, which makes conversational video avatars, introduced Griffin as a new model class for live face-to-face conversation: it perceives video and audio continuously and generates a talking video response in real time, handling overlap, interruptions and visual context (e.g. coaching a Rubik's cube solve). The headline claim comes from Tavus's own study; the VideoFDB numbers come from NVIDIA's benchmark leaderboard.

#### Why it matters
Real-time video avatars that half of short-call participants mistake for people move deepfake-style capability into live, interactive settings: customer service and tutoring, but also fraud and impersonation on video calls. The evidence so far is vendor-reported and based on one-minute calls.

Sources: [Tavus on X: launch-thread posts on full-duplex conversation (~707k and ~329k views)](https://x.com/tavus/status/2105704243890164166); [Tavus: Griffin, the first Human Interaction Model](https://www.tavus.io/griffin); [Business Wire (via Yahoo Finance): Tavus introduces Griffin](https://finance.yahoo.com/technology/ai/articles/tavus-introduces-griffin-first-face-163000466.html); [Tavus on X: launch post (~5.4M views)](https://x.com/tavus/status/2105704169009246248); [tech-ish: Nearly half of people on a video call with Griffin thought they were talking to a human](https://tech-ish.com/2026/10/02/tavus-griffin-ai-fooled-48-percent-video-call/); [Crypto Briefing: Tavus unveils Griffin, claiming the first video Turing test pass](https://cryptobriefing.com/tavus-griffin-video-turing-test/); [Kingy AI: Tavus Griffin specs, benchmarks, costs and availability](https://kingy.ai/blog/tavus-griffin-specs-benchmarks-pricing/); [Hassaan Raza (Tavus CEO) on X (~3.35M views)](https://x.com/hassaanraza/status/2105709935074185546); [Tavus on X: safety considerations before public release](https://x.com/tavus/status/2105704984914559441); [Kevin Roose on X (~400k views)](https://x.com/kevinroose/status/2105822736232911266); [Emad Mostaque on X (~570k views)](https://x.com/EMostaque/status/2105744708299526243); [Matt Walsh on X (~742k views)](https://x.com/MattWalshBlog/status/2106075822515773655); [@IterIntellectus on X (~906k views)](https://x.com/IterIntellectus/status/2105856072309449184); [Sen. Bernie Sanders on X: 'PAUSE AI NOW'](https://x.com/SenSanders/status/2106123796885430670); [Hassaan Raza on X: follow-up on safeguards and disclosure (Oct 3)](https://x.com/hassaanraza/status/2106185447361896475)

### 2026-10-02: Suno launches Speech (public beta): spoken voice plus background music generated in one track

**; Suno; Media generation; confidence medium; https://postcutoff.com/e/2026-10-02-suno-speech-beta/

Suno, the AI music company, released Speech in public beta for all web and mobile users. It is its first product beyond songs: it generates spoken narration with optional background music as a single track, from a short prompt (Simple mode) or a user script with voice gender, speech style and variety settings (Advanced mode). Suno warns that accents can drift and pauses can run long.

- Two modes in the Create tab: Simple (one-line description, e.g. 'a pirate captain rallying his crew') and Advanced (own script; voice gender, speech style, variation)
- Public beta for all users on web and mobile (The Verge)
- Known issues per Suno: accent drift (e.g. British drifting to Australian) and longer-than-requested pauses
- Launched while Suno's music product still faces copyright lawsuits (The Verge via Techmeme)

#### What happened
Suno extended its generator from songs to speech, positioning it for voiceovers, meditations, stories and dramatic readings with a score.

#### Why it matters
Music-generation companies are moving into the voice/TTS market led by ElevenLabs. Combined speech and score in one pass is a new
format for creators.

Sources: [The Verge: Suno's new Speech feature](https://www.theverge.com/ai-artificial-intelligence/1003925/suno-speech-ai-voice-feature-beta-availability); [Techmeme summary](https://www.techmeme.com/261002/p11); [AlphaSignal: Suno Speech generates narration and music together in one pass](https://alphasignal.ai/news/suno-speech-generates-narration-and-music-together-in-one-pass)

### 2026-10-04: Instagram adds a Meta AI song creator to DMs: type lyrics, pick a genre, send an AI-generated song

**; Meta; Media generation; confidence medium; https://postcutoff.com/e/2026-10-04-instagram-ai-song-creator-dms/

By early October 2026 Instagram had added a "create a song" option to its DM sharing menu. Users enter lyrics and choose a music genre, and Meta AI generates a track to send to friends. Social Media Today compared it to Snapchat's paid Snap-to-Song. It puts AI music generation in front of Instagram's very large user base.

- Spotted by creator Jonah Manzano on X (Sept 30): 'Instagram has added a new Meta AI song creator in DMs. Just type a few words or lyrics, choose a music style, and Meta AI can turn your idea into a song'
- Reported by Social Media Today on Oct 4–5, 2026; Meta called it a fun engagement option. The underlying music model is not named
- Comparable feature: Snapchat's Snap-to-Song for paying users, which turns message text into song lyrics (per Social Media Today; not otherwise verified here)

#### What happened
Instagram quietly rolled out AI song generation inside direct messages. It was first shown by a social-media tips creator, then
covered by trade press. The rollout date and the model behind it are not confirmed.

#### Why it matters
AI music generation moves from dedicated apps (Suno, Udio) into mainstream messaging, as label lawsuits and licensing deals over AI music
are still being settled.

Sources: [Social Media Today: Instagram adds AI songs to DMs](https://www.socialmediatoday.com/news/instagram-adds-ai-songs-to-dms/832086/); [Jonah Manzano on X: Instagram DMs can now turn your lyrics into songs](https://x.com/jonah_manzano/status/2105196063007334658)

### 2026-10-06: Google releases Nano Banana 2.1 image model (based on Gemini 3.6 Flash) at about half the per-image price of Nano Banana 2

***; Google DeepMind; Media generation; confidence high; https://postcutoff.com/e/2026-10-06-nano-banana-2-1/

On 6 Oct 2026 Google made Nano Banana 2.1 (gemini-nano-banana-2.1) generally available in the Gemini API, AI Studio and the Gemini app. Its model card says it is based on Gemini 3.6 Flash. Image output costs $30 per 1M tokens ($0.0336 per 1K image), half of Nano Banana 2's $60. Google deprecated the previous gemini-3.1-flash-image.

- API id gemini-nano-banana-2.1, GA on 6 Oct 2026 (Gemini API changelog), stable id only
- Pricing: $1.50 per 1M input tokens, $7.50 text+thinking output, $30 per 1M image-output tokens = $0.0336 (1K), $0.0504 (2K), $0.113 (4K) per image; Batch half price; no free tier (Google pricing page; The Decoder gives a different 4K figure)
- Model card: 'Nano Banana 2.1 is based on Gemini 3.6 Flash'; knowledge cutoff March 2026
- Thinking levels minimal/medium/high, grounding with Google Web and Image Search, up to 14 reference images; fixes tiling artifacts in extreme panorama ratios
- Rolled out in the Gemini app, AI Mode in Search, Flow, Stitch, Google Ads and Gemini Enterprise (The Decoder)
- X launch (16:00 UTC, views at 11:31 UTC on 7 Oct): Google ~901k, Google AI Studio ~479k, @NanoBanana ~163k; all say it 'outperforms our previous models across the board'. Google DeepMind's Philipp Schmid: '$0.034 per image vs $0.134', i.e. against Nano Banana Pro's 1K/2K price ('4x cheaper'), not Nano Banana 2's $0.067. MISMATCH, not resolved: he says 'up to 5 characters consistency', while the Gemini API model page says 4 characters (+10 objects) within 14 reference images

#### What happened
Google shipped the successor to Nano Banana 2 straight to general availability. It cut the image-output price in half and
built the model on its newer Gemini 3.6 Flash.

Google's launch posts on X (about 1.5M views together by 7 Oct) claim it "outperforms our previous models across the board",
with gains in visual design, mask-based editing and subject consistency. Philipp Schmid of Google DeepMind compared it with
Nano Banana Pro: "$0.034 per image vs $0.134", about four times cheaper. His "up to 5 characters" does not match the
official model page, which says character consistency for up to 4 characters; we have not resolved which is right.

#### Why it matters
Nano Banana is Google's default fast image generation and editing model across its products and API. Halving the per-image
price puts pressure on competing image APIs.

Sources: [Gemini API model page: gemini-nano-banana-2.1](https://ai.google.dev/gemini-api/docs/models/gemini-nano-banana-2.1); [Gemini API pricing](https://ai.google.dev/gemini-api/docs/pricing); [Gemini API changelog](https://ai.google.dev/gemini-api/docs/changelog); [Model card: Nano Banana 2.1](https://deepmind.google/models/model-cards/nano-banana-2-1/); [The Decoder: Google's new image model Nano Banana 2.1 generates better images for less money](https://the-decoder.com/googles-new-image-model-nano-banana-2-1-generates-better-images-for-less-money/); [Google on X: Meet Nano Banana 2.1 (~901k views)](https://x.com/Google/status/2107501209154204148); [Google AI Studio on X: introducing Nano Banana 2.1 (~479k views)](https://x.com/GoogleAIStudio/status/2107501303890915550); [@NanoBanana on X: Nano Banana 2.1 is here (~163k views)](https://x.com/NanoBanana/status/2107521202466046453); [Logan Kilpatrick on X: Nano Banana 2.1 is here (~110k views on 8 Oct)](https://x.com/OfficialLoganK/status/2107533224453058731); [Arena on X: Nano Banana 2.1 added to Text-to-Image, Image Edit and Multi-Image Edit Arenas (~66k views)](https://x.com/arena/status/2107576893973344578); [Philipp Schmid (Google DeepMind) on X: API id, $0.034 per image (~30k views)](https://x.com/_philschmid/status/2107502894685749672)

### 2026-10-07: Nieman Lab: ChatGPT images carry real New Yorker cartoonists' signatures, including Brendan Loper's 'BLOPER'

**; OpenAI; Media generation; confidence medium; https://postcutoff.com/e/2026-10-07-chatgpt-forges-new-yorker-cartoonist-signatures/

In early October 2026 Nieman Lab reported that ChatGPT's image generator puts the signatures or pen names of real New Yorker cartoonists on "New Yorker-style" cartoons. A cartoon signed "BLOPER" went viral on Facebook, and people asked Brendan Loper whether he had drawn it. Nieman Lab found more than 15 affected cartoonists, including Harry Bliss, Emily Flake, Joe Dator and Pat Byrnes. OpenAI's 2024 Condé Nast deal did not cover training on cartoons.

- Trigger: a ChatGPT image made for a Dolly Parton fan who asked for 'a New Yorker-style cartoon' carried 'BLOPER', the pen name of New Yorker cartoonist Brendan Loper, and spread on Facebook (Nieman Lab via AI Weekly/YourStory; Futurism, Oct 7)
- Scale: Nieman Lab found ChatGPT-generated images with the signatures of more than 15 New Yorker cartoonists (named: Harry Bliss, Emily Flake, Joe Dator, Pat Byrnes, Liza Donnelly; Futurism also lists Peter Vey and Jason Adam Katzenstein). A forged Loper signature had also been reported in May 2026
- Loper: 'I'm not a territorial person, but my name is my name'; 'It felt very much like a violation of my personhood'; his signature is 'the certificate of authenticity' (Futurism)
- OpenAI's 2024 licensing deal with Condé Nast did not permit training on cartoons; after being alerted, ChatGPT returned a guardrail message about 'similarity to third-party content' for such prompts (per summaries of Nieman Lab). OpenAI gave no public statement in Futurism's article

#### What happened
Image models reproduce visual conventions of their training data, including artists' signatures. Nieman Lab's investigation showed
that ChatGPT does this with the signatures of named, living New Yorker cartoonists, which can make readers believe an AI image is the
artist's work.

#### Why it matters
It is a concrete case of AI output being attributed to a real person through a forged mark, echoing Getty Images' claim against
Stability AI over reproduced watermarks, and it bears on whether image models trained on licensed or scraped work can avoid
reproducing identity markers.

Note: the Nieman Lab original was not located; facts come from Futurism and secondary summaries.

Sources: [Nieman Lab (original investigation): ChatGPT is adding real cartoonists' signatures to fake New Yorker cartoons](https://www.niemanlab.org/2026/10/chatgpt-is-adding-real-cartoonists-signatures-to-fake-new-yorker-cartoons/); [Futurism: ChatGPT caught putting signatures of real cartoonists on AI-generated comics (Oct 7)](https://futurism.com/artificial-intelligence/chatgpt-caught-putting-signatures-of-real-cartoonists-on-ai-generated-comics); [AI Weekly: ChatGPT signs AI 'New Yorker' cartoons with real artists' names (Nieman Lab summary)](https://aiweekly.co/alerts/chatgpt-signs-ai-new-yorker-cartoons-with-real-artists-names); [YourStory: ChatGPT is putting real artists' names on AI cartoons](https://yourstory.com/ai-story/chatgpt-real-artists-names-ai-cartoons)

### 2026-10-10: ShengShu releases Vidu Q4 Preview: 4K, 10-bit video up to 16 s with 15 reference images and 3 voice references, at a launch price from RMB 0.09 per second

**; ShengShu Technology; Media generation; confidence medium; https://postcutoff.com/e/2026-10-10-shengshu-vidu-q4-preview/

Around Oct 8-10, 2026 Beijing-based ShengShu Technology released Vidu Q4 Preview, the first preview of its next flagship video model. It outputs up to 4K with 10-bit colour and up to 16 seconds per generation, takes up to 15 reference images and 3 reference audio clips for consistent characters and voices, and costs from RMB 0.09 per second on Vidu's own app for a two-month launch period, below the RMB 0.5-3 per second typical of Chinese video APIs.

- Reported by IT Home on Oct 10, 2026; a hands-on review appeared on Tencent News on Oct 8 (per Pandaily, Oct 10)
- Output: 540p, 720p, 1080p, 2K and 4K, 10-bit colour, up to 16 seconds per generation; image-to-video and reference-to-video only for now
- References: up to 15 reference images (characters, scenes, props) and up to 3 reference audio clips to keep a character's voice consistent
- Price: two-month launch offer from RMB 0.09 per second (a 10-second clip under RMB 1) on Vidu's SaaS product; the standard price was not given
- Positioning: AI short dramas, advertising, social content and film production; the full Q4 release date was not announced
- No benchmark numbers; quality claims come from ShengShu and one reviewer

#### What happened
ShengShu, the Tsinghua-linked maker of the Vidu video models, previewed its Q4 generation. It focuses on consistent characters (many image
and voice references) and on a much lower price per second. We found no official ShengShu post in English; details come from Pandaily's
summary of IT Home and a Tencent News review.

#### Why it matters
Chinese video-model makers are competing hard on price. A launch price several times below typical API rates makes many takes affordable for
short-drama and ad production, which is where much AI video work in China happens.

Sources: [Pandaily: ShengShu releases Vidu Q4 Preview with 4K output and launch pricing from RMB 0.09 per second (Oct 10)](https://pandaily.com/shengshu-vidu-q4-preview-4k-video-model-reference-audio-0-09-yuan-per-second); [Vidu (product site)](https://www.vidu.com)

### 2026-10-14: YouTube Music starts labelling AI-made tracks and Deezer builds a filter to hide them, as over half of Deezer's daily uploads are AI-generated

***; YouTube, Google, Deezer; Media generation; confidence high; https://postcutoff.com/e/2026-10-14-youtube-music-ai-labels-deezer-filter/

On Oct 14, 2026 YouTube Music began showing labels on tracks made with generative AI, in three tiers declared by the distributor ("Fully Gen AI", "Partly Gen AI", "No Gen AI"), shown on the player page and coming to the main YouTube app; AI use alone does not block monetisation. The same day TechCrunch reported that Deezer will ship a user toggle, expected before the end of 2026, that hides AI-generated tracks, albums and artists from search. Deezer says AI tracks went from 10% of daily uploads in January 2025 to more than half, about 90,000 a day, by July 2026.

- YouTube Music labels appear on the player page and can be tapped for detail on what AI did; they will extend to the main YouTube app (TechCrunch)
- The three declarations partners choose from: 'Fully Gen AI' (entirely AI-generated), 'Partly Gen AI' (AI components plus the artist's own performance), 'No Gen AI' (includes pitch-correction tools such as Auto-Tune)
- YouTube says it combined partner declarations with its own detection tools; disclosure, not restriction: AI involvement alone does not make a track ineligible for monetisation
- Deezer: a toggle expected before the end of 2026 hides AI-generated tracks, albums and related artists from search results. CEO Alexis Lanternier: 'We've seen clear user demand for the ability to filter out AI-generated music'; a 2025 Deezer/Ipsos survey found about 45% of customers wanted AI music removed from their feeds
- Deezer's upload numbers: 10% of daily uploads when its detector launched in January 2025; more than 50% by July 2026, averaging about 90,000 fully AI-generated tracks a day. Deezer already labels detected AI music, keeps it out of algorithmic recommendations and editorial playlists, and withholds royalties from fraudulent streams
- Other platforms (TechCrunch): Apple Music plans 'Made With AI' labels as part of AI Transparency Tags this year; Spotify labels AI tracks and runs a spam filter but has not banned them; TIDAL labels fully AI-generated recordings, lets listeners turn them off and excludes them from royalty attribution; Bandcamp bans AI tracks

#### What happened
The two biggest streaming responses to AI music arrived on the same day: disclosure labels at YouTube and, at Deezer, the
option to not see AI music at all. YouTube's labels rest largely on what the uploader declares, with its own detection as a
check; Deezer's filter rests on its own detector, whose method it has not published.

#### Why it matters
AI tracks are now the majority of what Deezer receives each day, so the question has moved from whether to label to whether
listeners can opt out. The split between YouTube's disclose-and-monetise approach and Deezer's hide-it toggle sets the terms
for how AI-made music is treated commercially. The Verge's piece was not read beyond its headline.

Sources: [TechCrunch: YouTube Music launches new labels as AI-generated music floods streaming platforms (Oct 14)](https://techcrunch.com/2026/10/14/youtube-music-launches-new-labels-as-ai-generated-music-floods-streaming-platforms/); [TechCrunch: Deezer will soon let you filter out AI-generated music (Oct 14)](https://techcrunch.com/2026/10/14/deezer-will-soon-let-you-filter-out-ai-generated-music/); [The Verge: YouTube sticks new labels on AI-generated songs (Oct 14)](https://www.theverge.com/ai-artificial-intelligence/1011369/youtube-music-ai-label)

### 2026-10-15: Black Forest Labs releases FLUX 3 Image Klein, a 5.6B open-weights image generation and editing model

***; Black Forest Labs; Media generation; confidence high; https://postcutoff.com/e/2026-10-15-flux-3-image-klein-open-weights/

On Oct 15, 2026 Black Forest Labs released FLUX 3 Image Klein, a 5.6B-parameter open-weights model for text-to-image, image editing and multi-reference generation (up to 10 reference images) with native 4K output and bounding-box control. BFL's own Elo tests put it first among open-weights image models and above closed Ideogram V4.5 and FLUX.2 [pro]. Weights are under the FLUX Kommunity License (free commercial use under $5M annual revenue); the API price is $0.009 per 1K image.

- Size: 5.6B image transformer (text encoder Qwen3-VL-4B-Instruct, Apache-2.0, separate); base checkpoint (50 steps, guidance 3.5) for fine-tuning and a step-distilled checkpoint (16 steps, no guidance) for low latency; FP8, INT8 and NVFP4 builds
- Capabilities shared with FLUX 3 Image: native 4K generation, bounding boxes (0–1000 grid) to place, move, recolor, replace or remove elements; one model for text-to-image, single-reference editing and up to 10 reference images
- BFL's internal Elo: first among open-weights models in text-to-image, multi-reference and editing; above closed Ideogram V4.5 and FLUX.2 [pro]; vs FLUX.2 [klein] 9B +91 Elo (text-to-image) and +70 (editing) with 38% fewer parameters; compared with Qwen-Image 2.1 (7B) and Krea 2 (12.9B)
- Speed (BFL): API median 4.54 s (text-to-image) and 7.18 s (editing) at 1 MP incl. queueing; locally ~4 s per 1 MP on an RTX 5090 and under 9 s on an RTX 4070 (distilled, INT8, ComfyUI); under 0.6 s on a GB300 with NVFP4; INT8 build runs on 8 GB VRAM
- Model card: 'Do not use the model without a prompt upsampler' (flux-3-image-klein-pup, a 9B model); peak memory on a B200 at 1024² without upsampler 31 GB (BF16) / 26 GB (FP8) / 25 GB (NVFP4)
- Licence: FLUX Kommunity License v1.0: commercial output generation for users under US$5M gross annualized revenue; otherwise a commercial licence from BFL; Hugging Face access gated (contact details)
- API: flux-3-image-klein on the BFL API; per-image price $0.008 (768²), $0.009 (1K), $0.011 (1.5K), $0.014 (2K), $0.036 (4K), same for generation and editing

#### What happened
Black Forest Labs (BFL) published FLUX 3 Image Klein on Oct 15, 2026 (the Hugging Face repository was created on Oct 13
and was trending on Oct 16). It is the small, open-weights member of the FLUX 3 family announced in July. The 5.6B model
handles text-to-image, editing and multi-reference generation in one network and keeps two features of the larger FLUX 3
Image: native 4K output and bounding boxes that steer where elements go (they guide placement but are not hard masks).

BFL ships a base checkpoint for fine-tuning and LoRA training and a step-distilled one for fast generation, plus FP8, INT8
and NVFP4 builds. The model card insists on a separate 9B prompt upsampler, because short prompts give generic results. The
blog says the INT8 build runs on GPUs with 8 GB of VRAM; the card's B200 memory table (25–31 GB without the upsampler)
measures a different setup, so local requirements depend on the build and the pipeline.

All quality and speed numbers are BFL's own (internal Elo evaluations; API latency measured end to end, other models timed
from the client, so BFL itself says the comparison is not like for like).

#### Why it matters
It brings FLUX 3-class editing and 4K output to consumer GPUs a day after Krea 3 and a month after Qwen-Image 2.1, and it
uses the same "free under $5M revenue" licence pattern that open image models increasingly adopt instead of Apache-2.0
(FLUX.2 [klein] 4B was Apache-2.0).

Sources: [Black Forest Labs: FLUX 3 Image Klein (blog)](https://bfl.ai/blog/flux-3-image-klein); [Hugging Face: black-forest-labs/flux-3-image-klein (model card)](https://huggingface.co/black-forest-labs/flux-3-image-klein); [BFL API pricing](https://docs.bfl.ai/quick_start/pricing); [GitHub: reference code](https://github.com/black-forest-labs/flux-3-image-klein)

### 2026-10-15: Krea releases Krea 3, an image-editing model it says is 4× faster and about half the price of Nano Banana Pro (open beta)

**; Krea; Media generation; confidence medium; https://postcutoff.com/e/2026-10-15-krea-3-image-editing-model/

On Oct 15, 2026 Krea opened the beta of Krea 3, its own foundation image model aimed at production image editing. Krea says it renders only the region that changes, finishes a 2K edit in under 8 seconds and costs about $0.06 per edit, against ~26.8 s and ~$0.135 for Google's Nano Banana Pro and ~39.9 s and ~$0.21 for GPT Image 2.5 in its own comparison. Krea's speed, price and quality claims are the company's own; no independent benchmark yet.

- Status: open beta announced Oct 15, 2026 by Krea co-founder Diego (HN user dvrp); available in the Krea app and the Krea API, with enterprise self-hosting 'on your own GPUs in your cloud'
- Krea's comparison (one image, twelve identical edit prompts, 2K, median request-to-result time including queue): Krea 3 under 8 s and ~$0.06 per edit; Nano Banana Pro ~26.8 s, ~$0.135; GPT Image 2.5 ~39.9 s, ~$0.21 (Krea calls the GPT Image 2.5 price a draft estimate)
- Design claim: Krea 3 'finds that region first and renders only what changes'; Krea says it is 'not a distilled copy of a larger model'
- Features Krea lists: native regional edits, pixel-precise edits (a 'Precise mode' that leaves pixels outside the edit area unchanged), plugins for Figma, Photoshop, After Effects, Premiere Pro, DaVinci Resolve and Blender
- Pricing: 'The launch price is the list price. We expect it to go down, not up.'
- Future (founder on HN): a technical report like Krea 2's is planned; LoRA fine-tuning and open weights are being considered, not promised
- Hacker News: 172 points on launch day; commenters asked why FLUX 3 was not compared and criticised the sign-up discount flow

#### What happened
Krea, the AI image and video design platform, announced the open beta of Krea 3 on Oct 15, 2026. It is the company's own
foundation image model and is built for editing rather than for generation from scratch. The launch page compares it with
Google's Nano Banana Pro and OpenAI's GPT Image 2.5 on the same image and twelve identical prompts at 2K resolution. On that
page Krea 3 takes under 8 seconds and about $0.06 per edit, against about 26.8 s and $0.135 for Nano Banana Pro and 39.9 s
and $0.21 for GPT Image 2.5. Krea says the model first finds the region to change and renders only that region. A
"precision" demo counts the pixels outside the edit area that each model leaves unchanged.

In the Hacker News launch thread, co-founder Diego (dvrp) said benchmarks are "still being worked on". He said a technical
report will follow, as it did for Krea 2, and that the team is thinking about LoRA fine-tuning and open weights.

#### Why it matters
Image-editing models are moving from quality races to cost and latency. Krea is a large creative-tools platform that until
now mostly resold other labs' models, and it is now pricing its own model below the frontier labs' editors. The comparison is
the vendor's own and is not yet independently checked.

Sources: [Krea: Krea 3 — frontier image editing model](https://www.krea.ai/krea-3); [Hacker News discussion (founder's launch comment)](https://news.ycombinator.com/item?id=50105648); [Krea 2 technical report (earlier model)](https://www.krea.ai/blog/krea-2-technical-report)

### 2026-10-17: Three unannounced OpenAI 'aurora' models briefly appear in the API and then in Arena image tests, possibly GPT-Image-3 (leak, unconfirmed)

**; OpenAI; Media generation; confidence medium; https://postcutoff.com/e/2026-10-17-openai-aurora-image-models-arena-leak/

Over the weekend of Oct 17–18, 2026, references to three models, aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha, briefly appeared in OpenAI's API and were removed; within hours they showed up in blind image-generation battles on Arena. TestingCatalog and Lumina Bench guess they are a next OpenAI image model (GPT-Image-3). OpenAI has not commented, and there are no docs or prices.

- Names: aurora-red-alpha, aurora-green-alpha, aurora-blue-alpha; first spotted via an alert on the Dev Mode Discord server, then removed from the API (TestingCatalog suggests an accidental rollout)
- Arena: Lumina Bench (@LuminaBench) reported all three as image models in blind tests; red was faster but lower quality, green slower and better; blue untested at the time
- Attribution to GPT-Image-3 is a guess by TestingCatalog and Lumina Bench, not confirmed by OpenAI
- Separate, unlinked: three new Design Arena entries named Aspen, Cypress and Maple were reported by @kdnbmb about an hour later
- Previous OpenAI image release in the dataset period: ChatGPT Images 2.5 (API: GPT-Image-2.5 Flare for speed, Sunburst for precision), per TestingCatalog's background

#### What happened
Model identifiers for three "aurora" variants were visible in OpenAI's API for a short time and then removed. Hours later the same
names appeared on Arena (arena.ai), where testers compare image models blind. Early samples (cat portraits) were judged clean and
detailed but not striking. Nothing ties the names to a product yet.

#### Why it matters
Arena appearances have preceded most recent OpenAI image releases, so this is a likely early sign of a new image model. It is a
leak: treat the GPT-Image-3 label as speculation until OpenAI announces something. Exact date of the API sighting (Oct 17 or 18)
is not given in the article; the entry uses Oct 17.

Sources: [TestingCatalog: OpenAI possibly testing GPT-Image-3 ('Aurora' models on Arena), Oct 18](https://www.testingcatalog.com/openai-possibly-testing-gpt-image-3/)

### 2026-10-19: fal and Black Forest Labs release FLUX 3 Max, a post-trained video model that makes a 10-second 720p clip in about 12 seconds

**; fal, Black Forest Labs; Media generation; confidence high; https://postcutoff.com/e/2026-10-19-flux-3-max-fal/

On Oct 19, 2026 inference provider fal announced FLUX 3 Max, a video model its research team post-trained on Black Forest Labs' FLUX 3 base. fal says it renders a 10-second 720p clip in about 12.5 s, against about 126 s for FLUX 3 and 278 s for Seedance 2.5, while tying for #1 on Artificial Analysis' preliminary image-to-video arena. Launch price $0.0425 per second (75% off until Nov 7).

- Built by fal Research with Black Forest Labs on the FLUX 3 base model; text-to-video, image-to-video and keyframes-to-video on fal's API
- Speed (fal): ~12.45 s for a 10 s 720p video vs 126.44 s for FLUX 3 and 277.80 s for Seedance 2.5
- Quality (fal): Elo 1139 on Artificial Analysis' preliminary image-to-video arena, tied #1 with Seedance 2.5 (MiniMax H3 1141); AA text-to-video board read 2026-10-25: #6, Elo 1155
- Price: $0.0425 per second at 720p, discounted 75% from $0.17 until Nov 7, 2026
- fal's X announcement (Oct 19) had ~444k views by Oct 25

#### What happened
fal, a model-hosting company, released a faster post-trained version of Black Forest Labs' FLUX 3 video model, served only through
its own platform, with a large launch discount.

#### Why it matters
Generation time for a 10-second clip falls to roughly real time, which matters for interactive and agent-driven video tools.

Sources: [fal: FLUX 3 Max](https://fal.ai/flux-3-max); [fal on X: FLUX 3 Max announcement (Oct 19)](https://x.com/fal/status/2112251022047412492)

## Models

### Lyria 3 (Clip / Pro) (Google DeepMind, 2026-02-18, legacy)

Type music. https://postcutoff.com/m/lyria-3/
Access: Gemini API (Interactions API) `lyria-3-clip-preview`; Gemini API (Interactions API) `lyria-3-pro-preview`; Google Cloud Gemini Enterprise Agent Platform (Vertex AI) `lyria-3-pro-preview`; Google Cloud Gemini Enterprise Agent Platform (Vertex AI) `lyria-3-clip-preview`; OpenRouter `google/lyria-3-pro-preview`; Web app https://gemini.google.com
- Songs with vocals and auto-written lyrics in the Gemini app: 30-second tracks with vocals and lyrics from a text prompt, photo or video, with Nano Banana cover art; 8 languages (en, de, es, fr, hi, ja, ko, pt); 18+ only.
- SynthID watermark + detection in Gemini: All outputs carry SynthID; the Gemini app can check whether uploaded audio was generated with Google AI via SynthID.
- Full songs with structure control (Lyria 3 Pro): Tracks up to ~3 minutes (184 s max on Vertex) with control over intros, verses, choruses and bridges, duration, BPM and intensity; 44.1 kHz, 192 kbps MP3; C2PA content credentials and vocal-likeness filtering.

### Suno v5.5 (Suno, 2026-03-26, retired)

Type music. https://postcutoff.com/m/suno-v5-5/
Access: Web app https://suno.com
- Voices (sing with your own voice): Record/upload your voice (with verification and privacy controls) and have Suno sing songs in it; Pro/Premier.
- Custom Models: Fine-tune a personal v5.5 on your own catalog (min. 6 tracks, up to 3 models per user); Pro/Premier.
- My Taste personalization: Learns preferred genres/moods and applies them via the Magic Wand; all users.

### Eleven Dubbing v2 (ElevenLabs, 2026-05-28, preview)

Type audio/speech. https://postcutoff.com/m/elevenlabs-dubbing-v2/
Access: ElevenLabs API; Web app (ElevenCreative / ElevenProductions) https://elevenlabs.io/dubbing
- Direct speech-to-speech dubbing: Conditions directly on the original performance instead of an ASR -> translate -> TTS pipeline, so intonation and emotion carry across 90+ languages; ElevenLabs: 'For the first time, the emotion and performance of the original speaker carries across every language' (company claim, not independently verified as a first).
- Project-based dubbing API: API (2026-08-06/10) with editable JSON transcripts/translations, regional variants (e.g. es-MX), sync-aware translation; 3 GB per file via API.

### Lyria 3.5 (Google DeepMind, 2026-07-29, current)

Type music. https://postcutoff.com/m/lyria-3-5/
Access: Gemini API (Interactions API) `lyria-3.5`; Gemini API (Interactions API) `lyria-3-clip-preview`; Google Cloud Gemini Enterprise Agent Platform (Vertex AI); OpenRouter `google/lyria-3-pro-preview`; Web app https://gemini.google.com
- Full songs with vocals and lyrics: Full-length ~2-minute tracks with verses/choruses/bridges, generated vocals and lyrics; 44.1 kHz stereo MP3/WAV.
- Image-conditioned music: Accepts text and image prompts via the Interactions API.
- SynthID-watermarked audio: Latent-diffusion model with SynthID watermarking on outputs.

### Dreamina Seedance 2.5 (ByteDance Seed, 2026-07-31, current)

Type video-gen. https://postcutoff.com/m/seedance-2-5/
Access: BytePlus ModelArk `dreamina-seedance-2-5-260628`; OpenRouter `bytedance/seedance-2.5`; Web app (Jimeng / Doubao Pro, China) https://seed.bytedance.com/en/seedance2_5; Web app (Dreamina, CapCut) https://dreamina.capcut.com/seedance/seedance-2-5
- 30-second single-shot generation with extensions: Up to 30 s of joint audio-video per generation (15 s in Seedance 2.0), extendable twice with consistent characters, settings and pacing.
- 50 multimodal references in one pass: Up to 30 images, 10 video clips and 10 audio clips as references in a single request.
- Timestamp-level audio/video editing: Targeted edits of audio and video at given timestamps, plus green-screen, white-model (untextured previs) control and camera-perspective editing.
- Draft-then-final workflow: API can render a cheap 480p Draft first and then generate the final video from the Draft task ID (final currently 1080p only; Draft IDs valid 7 days).

### LTX-2.5 (Lightricks, 2026-08-11, current)

Type video-gen, open weights. https://postcutoff.com/m/ltx-2-5/
Access: Hugging Face https://huggingface.co/Lightricks/LTX-2.5; GitHub https://github.com/Lightricks/LTX-2; LTX API https://console.ltx.io/playground/
- Open audio-video generation with multishot consistency: 22B DiT that generates synchronized video and audio and keeps character, environment, lighting and voice consistent across cuts.
- Fast generation: 10-second 720p image-to-video clip in 6.8 s on two GB200s (vendor claim); distilled variant uses 8 steps.
- Robotics / physical-AI checkpoint: Ships a pretrained checkpoint aimed at world-model use in robotics and physical AI.

### Wan 3.0 (Alibaba Cloud (Tongyi Lab), 2026-08-24, current)

Type video-gen. https://postcutoff.com/m/wan-3-0/
Access: Alibaba Cloud Model Studio (DashScope) `wan3.0-video`; Alibaba Cloud Model Studio (faster Prime tier) `wan3.0-video-prime`; Qwen Cloud `wan3.0-video` https://www.qwencloud.com/models/wan3.0-video; OpenRouter `alibaba/wan-3.0`; OpenRouter (Prime) `alibaba/wan-3.0-prime`
- 30-second single-pass clips with native audio: Up to 30 s per generation (2-30 s or smart duration) at up to 1080P, with dialogue, music and sound effects generated together; double Wan 2.7's 15 s limit.
- Documents and web pages as input: Besides text, images, video and audio, one request can take a file (pdf, docx, pptx, xlsx, txt, md) or a public web link as source material for the video.
- Multi-reference generation and editing: Up to 20 references per request (10 images, 5 videos, 5 audio clips, 15 s total each for video/audio), plus first/last-frame modes and editing of an existing generation.

### MAI-Image-2.6-Flash (Microsoft, 2026-09-04, preview)

Type image-gen. https://postcutoff.com/m/mai-image-2-6-flash/
Access: Microsoft Foundry (Foundry Models sold by Azure) `MAI-Image-2.6-Flash`; Microsoft Foundry (image edits) `MAI-Image-2.6-Flash`; OpenRouter `microsoft/mai-image-2.6-flash`; Web app (MAI Playground) https://playground.microsoft.ai/
- Low-cost, fast tier of MAI-Image-2.6: Same generation and editing capabilities as MAI-Image-2.6 at lower cost; Microsoft claims 2.8x faster than GPT-Image-2-Medium with '72% greater efficiency' (metric not defined).
- Bing web grounding and auto aspect ratio: Supports web_grounding (Bing Search context) and auto_aspect_ratio, like MAI-Image-2.6; up to 2,359,296 px output (1536x1536 equivalent).
- Up to five reference images for edits: Edits API accepts up to five JPEG/PNG references for multi-image composition and precise local edits.

### MAI-Image-2.6 (Microsoft, 2026-09-04, preview)

Type image-gen. https://postcutoff.com/m/mai-image-2-6/
Access: Microsoft Foundry (Foundry Models sold by Azure) `MAI-Image-2.6`; Microsoft Foundry (image edits) `MAI-Image-2.6`; OpenRouter `microsoft/mai-image-2.6` https://openrouter.ai/microsoft/mai-image-2.6; Web app (MAI Playground) https://playground.microsoft.ai/
- Bing web grounding: web_grounding: true lets the model retrieve current information from Bing Search and use it as context for real-world entities, places and events (MAI-Image-2.6 and -Flash only).
- Model-chosen aspect ratio: auto_aspect_ratio: true lets the model pick the output aspect ratio from the prompt and input images; output up to 2,359,296 px total (1536x1536 equivalent), minimum 768 px per side.
- Multi-image reference editing: Edits API takes up to five JPEG/PNG reference images and combines people, products, styles and scenes; precise local edits (object removal, inpainting, text updates) that preserve layout.

### GPT Image 2.5 Flare (OpenAI, 2026-09-08, current)

Type image-gen. https://postcutoff.com/m/gpt-image-2-5-flare/
Access: OpenAI API `gpt-image-2.5-flare`; Azure OpenAI (Microsoft Foundry) `gpt-image-2.5-flare`; Web app https://chatgpt.com; ElevenLabs Image & Video API `gpt-image-2.5-flare`
- Fast everyday image generation: Fastest high-quality OpenAI image model; quality low/medium/high/xhigh/max/auto.
- Inpainting: Editing with masks via v1/images/edits.

### GPT Image 2.5 Sunburst (OpenAI, 2026-09-08, current)

Type image-gen. https://postcutoff.com/m/gpt-image-2-5-sunburst/
Access: OpenAI API `gpt-image-2.5-sunburst`; Azure OpenAI (Microsoft Foundry) `gpt-image-2.5-sunburst`; Web app https://chatgpt.com; ElevenLabs Image & Video API `gpt-image-2.5-sunburst`
- Most capable OpenAI image model: Top-quality generation and editing with inpainting via images/generations and images/edits.
- Replacement for gpt-image-1.5/1-mini: Named successor for image models shutting down Dec 1 2026.

### Suno v6 (v6, v6-wild, v6-mini) (Suno, 2026-09-09, current)

Type music. https://postcutoff.com/m/suno-v6/
Access: Web app `v6` https://suno.com; Web app (Pro/Premier) `v6-wild`; Web app (all users, incl. free) `v6-mini`
- Trained only on licensed music: First Suno generation developed with rightsholders; trained from scratch on music licensed from Warner Music Group, BMG and Believe (not on data used for earlier Suno versions), with revenue sharing to partners.
- Three-variant lineup: v6 (reliable, steerable flagship), v6-wild (experimental, genre-blending, pushes away from the prompt), v6-mini (fast, high-volume, available to everyone).
- Natural-language section and lyric editing: Edit parts of a song or change individual lyric lines by prompt without regenerating the whole track.
- Multimodal references and mashups: Text, audio, image and video references as a starting point; combine elements of several songs into a mashup; sample/isolate instruments and build beats.
- Upload screening and download limits: Uploaded audio and lyrics are screened for unauthorized use; downloads are capped per plan (none on Free, 20/month Pro, 60/month Premier).

### Eleven Music v2.5 (ElevenLabs, 2026-09-11, current)

Type music. https://postcutoff.com/m/elevenlabs-music-v2-5/
Access: ElevenLabs API `music_v2_5`; Web app (ElevenMusic) https://elevenmusic.io
- Commercially cleared music generation: Richer melodies and live-sounding instruments, built for commercial use; lossless downloads on every plan incl. Free.
- Composition plans and audio reference: Music v2 line supports structured composition plans and reference-audio generation (v2.5 default for prompted and reference generation).
- Composition-plan chunks via API: API support rolled out 2026-09-14 with 6,132-character composition chunks; waveform visual data via with_waveform_visual (2026-08-03).

### Qwen-Image-2.1 (Alibaba (Qwen), 2026-09-20, current)

Type image-gen, open weights. https://postcutoff.com/m/qwen-image-2-1/
Access: Hugging Face `Qwen/Qwen-Image-2.1` https://huggingface.co/Qwen/Qwen-Image-2.1; ModelScope https://www.modelscope.cn/models/Qwen/Qwen-Image-2.1; Hugging Face (Turbo, 8-step) `Qwen/Qwen-Image-2.1-Turbo` https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo; GitHub https://github.com/QwenLM/Qwen-Image-2.1; Hugging Face Space (demo) https://huggingface.co/spaces/Qwen/Qwen-Image-2.1
- Native RGBA generation and editing: Generates transparent images, edits transparent layers and extracts subjects from photos in one model.
- Multi-reference editing: Up to 10 reference images; local edits via circles, painted annotations or masks, with identity preservation.

### HeyGen Video 1 (HeyGen, 2026-09-30, current)

Type video-gen. https://postcutoff.com/m/heygen-video-1/
Access: HeyGen API `heygen-video-1`; OpenRouter `heygen/heygen-video-1`; Web app https://app.heygen.com/apps/creative-studio/video
- Whole scene with sound in one pass: General-purpose (non-avatar) model: picture, dialogue, ambience and sound effects generated together; AAC 32 kHz stereo in MP4.
- Text, image and reference modes: text_to_video, image_to_video and reference_to_video (default), with up to 9 images, 3 videos and 3 audio references (12 total).
- Up to 30 s and 2K (added after launch): October 2026 updates added auto duration (8-30 s from quoted dialogue), 5-30 s for t2v/i2v with turbo or quality enhancement, and 1080p and 2K outputs (billed at 3x 768p).

### Utopai X (Utopai Studios, 2026-09-30, current)

Type video-gen. https://postcutoff.com/m/utopai-x/
Access: Web app (PAI) https://pai.utopaistudios.com/
- Production-aware agentic video model: Built into PAI and works from the whole production (scripts, world-building assets, production status, collaborator feedback) so each generated shot fits the film.
- Reflections, caustics and spatial consistency: Utopai says it is particularly strong at reflections, caustics, world understanding and high-level spatial consistency.

### Tavus Griffin (Human Interaction Model) (Tavus, 2026-10-01, preview)

Type video-gen. https://postcutoff.com/m/tavus-griffin/
Access: Tavus (research preview) https://www.tavus.io/griffin
- Full-duplex face-to-face video conversation: Video-to-video model that sees and hears the user while speaking, interrupts and takes interruptions, and reacts to gestures and objects in real time; Tavus calls it the first 'Human Interaction Model'.
- 'Video Turing test' result (vendor study): 48% of 54 participants in 1-minute blind live calls judged it human (Tavus's earlier stack ≤2%); NVIDIA VideoFDB generation 3.83/5 vs human 3.92; median latency ~1.9 s.

### Nano Banana 2.1 (Gemini Nano Banana 2.1) (Google DeepMind, 2026-10-06, current)

Type image-gen. https://postcutoff.com/m/gemini-nano-banana-2-1/
Access: Gemini API `gemini-nano-banana-2.1`; Google AI Studio https://aistudio.google.com?model=gemini-nano-banana-2.1; OpenRouter `google/gemini-nano-banana-2.1`; Web app https://gemini.google.com
- Half the per-image price of Nano Banana 2: Image output is $30 per 1M tokens versus $60 for gemini-3.1-flash-image: a 1K image costs $0.0336 instead of $0.067. Input is pricier ($1.50 vs $0.50 per 1M).
- Built on Gemini 3.6 Flash: Model card: 'Nano Banana 2.1 is based on Gemini 3.6 Flash', with a March 2026 knowledge cutoff.
- Configurable thinking and search grounding: Thinking levels minimal / medium (default) / high; grounding with Google Web and Image Search; up to 14 reference images (4 consistent characters, 10 objects).
- Fixed panorama tiling: Fixes tiling artifacts at 1:4, 4:1, 1:8 and 8:1 aspect ratios at 2K and 4K, and improves text and infographic layout.

### FLUX 3 Image Klein (Black Forest Labs, 2026-10-15, current)

Type image-gen, open weights. https://postcutoff.com/m/flux-3-image-klein/
Access: BFL API `flux-3-image-klein`; Hugging Face https://huggingface.co/black-forest-labs/flux-3-image-klein; GitHub (reference code) https://github.com/black-forest-labs/flux-3-image-klein
- Open-weights 4K generation with bounding boxes: 5.6B model with FLUX 3 Image's native 4K output and 0–1000-grid bounding boxes to place, move, recolor, replace or remove elements; up to 10 reference images.
- Consumer-GPU builds: Distilled INT8 build runs on 8 GB VRAM; ~4 s per 1 MP on an RTX 5090, under 9 s on an RTX 4070 (BFL figures, ComfyUI).

### Krea 3 (Fast / High / Ultra modes) (Krea, 2026-10-15, current)

Type image-gen. https://postcutoff.com/m/krea-3/
Access: Krea API `krea-3`; OpenRouter `krea/k3-edit`; Web app https://www.krea.ai/krea-3
- Ultra thinking mode: thinking_mode 'ultra' generates several candidates, reviews them against the prompt and references and returns the best at 2K; needs at least one input image and a Max plan; jobs can take up to ~10 minutes.
- Pixel-exact regional edits: regional_edit: true regenerates only the parts that must change and leaves the rest pixel-exact, so edits can be chained without quality loss.
- Bounding boxes with per-region notes: Up to 24 regions (0-1 coordinates) on any reference image, each with its own @label and note prompt; up to 8 reference images per request.
- Self-hosting on enterprise plans: Krea says Krea 3 can run on the customer's own GPUs in their cloud under enterprise terms.

### Kling 4.0 (and 4.0 Flash) (Kuaishou, 2026-10-15, current)

Type video-gen. https://postcutoff.com/m/kling-4-0/
Access: fal.ai `kling/v4/image-to-video`; fal.ai `kling/v4/text-to-video`; fal.ai `kling/v4/reference-to-video`; fal.ai (Omni video edit) `kling/v4/video-to-video/edit`; fal.ai (4.0 Flash) `kling/v4/flash/text-to-video`; Web app https://kling.ai
- 30-second single-pass clips: Native generation of up to 30 s in one pass (Kling 3.0: 15 s); fal's endpoints take 3-30 s, 720p or 1080p, with per-shot multi-prompts.
- Omni Reference with keyframes: Up to 15 reference assets and up to 10 keyframe inputs per generation (Kuaishou release); start and end frame on image-to-video.

### FLUX 3 Max (Black Forest Labs, 2026-10-19, current)

Type video-gen. https://postcutoff.com/m/flux-3-max/
Access: fal `blackforestlabs/flux-3-max/text-to-video`; fal `blackforestlabs/flux-3-max/image-to-video`; fal `blackforestlabs/flux-3-max/keyframes-to-video`
- Fast frontier-quality video: fal says a 10 s 720p clip takes about 12.45 s, versus 126.44 s for FLUX 3 and 277.80 s for Seedance 2.5 (fal's own measurement).
- Post-trained FLUX 3 Video by fal Research: Not a separate BFL tier: fal Research post-trained BFL's FLUX 3 Video with BFL for prompt adherence and lower latency; +50 Elo over FLUX 3 on AA image-to-video per fal.
- Keyframes-to-video: Up to ten images placed at chosen frame indexes, with a prompt describing the motion between them; 5-20 s, 720p or 1080p, audio on by default.

## Videos

- 2026-10-01, Tavus: [48% of People Thought This AI Was a Real Human | Introducing Griffin](https://www.youtube.com/watch?v=lHw6yoyPkpo) https://postcutoff.com/v/tavus-48-of-people-thought-this-ai-was-a-real-human-introducing-gr/
- 2026-09-28, Kling AI: [The Beat | Made with KLING 4.0](https://www.youtube.com/watch?v=w3397LF5MAc) https://postcutoff.com/v/kling-the-beat-made-with-kling-4-0/
- 2026-09-11, ElevenLabs: [Introducing Music v2.5](https://www.youtube.com/watch?v=zXlVQ8rMJM0) https://postcutoff.com/v/elevenlabs-introducing-music-v2-5/
- 2026-09-04, Google Gemini: [Introducing Lyria 3.5 Music Generation in Gemini](https://www.youtube.com/watch?v=WLRDC_aGCr8) https://postcutoff.com/v/google-gemini-introducing-lyria-3-5/
- 2026-05-24, Lucas Cinematic Studio: [What Remains | Short Film | Finalist · Seoul International AI Film Festival 2026](https://www.youtube.com/watch?v=EaTvBF1I6ZQ) https://postcutoff.com/v/lucas-kern-what-remains-seoul-ai-film-festival/
- 2026-05-19, Google for Developers: [Introducing Gemini Omni](https://www.youtube.com/watch?v=5T0yRNmNRi4) https://postcutoff.com/v/introducing-gemini-omni-google-for-developers/
- 2026-05-19, Google: [Introducing Gemini Omni: Create Anything from Anything](https://www.youtube.com/watch?v=KUyRq7szZsM) https://postcutoff.com/v/introducing-gemini-omni-google/
- 2026-02-28, Lennard Smith: [BONE THRONE | AI Short Film Made with Seedance 2.0 & Kling 3.0](https://www.youtube.com/watch?v=6D4_ZMnPx7I) https://postcutoff.com/v/lennard-smith-bone-throne/
- 2025-06-11, PJ Ace: [Disney approved our insane AI Kalshi ad to run during the NBA Finals 🤣](https://www.youtube.com/watch?v=-QMftwmyW-A) https://postcutoff.com/v/pj-ace-kalshi-nba-finals-veo-3-ad/
- 2025-02-13, OpenAI: [匚尺丨ㄒㄒ乇尺乙 — REMASTERED with Sora](https://www.youtube.com/watch?v=qjuk0YCUdo8) https://postcutoff.com/v/openai-critterz-remastered-with-sora/
- 2024-05-02, Washed Out: [Washed Out - The Hardest Part (Official Video)](https://www.youtube.com/watch?v=-Nb-M1GAOX8) https://postcutoff.com/v/washed-out-the-hardest-part-sora/
- 2024-04-05, OpenAI: [air head · Made by shy kids with Sora](https://www.youtube.com/watch?v=9oryIMNVtto) https://postcutoff.com/v/shy-kids-air-head-sora/
- 2024-02-28, Just A Happy Troll: [Will Smith Eating Spaghetti AI Video - (2023 vs 2024)](https://www.youtube.com/watch?v=vbWe5k4fFWE) https://postcutoff.com/v/will-smith-eating-spaghetti-2023-vs-2024/

## Posts

- 2026-10-19, fal: [](https://x.com/fal/status/2112251022047412492)
- 2026-10-06, Google: [Google: "Meet Nano Banana 2.1, our latest image generation and editing model"](https://x.com/Google/status/2107501209154204148)
- 2026-10-06, Arena.ai: [](https://x.com/arena/status/2107576893973344578)
- 2026-10-06, Logan Kilpatrick: [](https://x.com/OfficialLoganK/status/2107533224453058731)
- 2026-10-06, Google AI Studio: [Google AI Studio: "introducing @NanoBanana 2.1", try it in AI Studio](https://x.com/GoogleAIStudio/status/2107501303890915550)
- 2026-10-06, Nano Banana: [@NanoBanana: "Nano Banana 2.1 is here!"](https://x.com/NanoBanana/status/2107521202466046453)
- 2026-10-06, Philipp Schmid: [Philipp Schmid (Google DeepMind): gemini-nano-banana-2.1 at "$0.034 per image vs $0.134"](https://x.com/_philschmid/status/2107502894685749672)
- 2026-10-03, Hassaan Raza: [](https://x.com/hassaanraza/status/2106185447361896475)
- 2026-10-02, vittorio: [](https://x.com/IterIntellectus/status/2105856072309449184)
- 2026-10-02, Kevin Roose: [](https://x.com/kevinroose/status/2105822736232911266)
- 2026-10-02, Matt Walsh: [Matt Walsh: an AI video chat that passes for human is 'a disaster for the human race'](https://x.com/MattWalshBlog/status/2106075822515773655)
- 2026-10-02, Sen. Bernie Sanders: [Bernie Sanders on Tavus Griffin: 'PAUSE AI NOW'](https://x.com/SenSanders/status/2106123796885430670)
- 2026-10-01, Tavus: [Tavus: 'Introducing Griffin, the first model to pass the video Turing test'](https://x.com/tavus/status/2105704169009246248)
- 2026-10-01, Emad: [](https://x.com/EMostaque/status/2105744708299526243)
- 2026-10-01, Tavus: [](https://x.com/tavus/status/2105704243890164166)
- 2026-10-01, Tavus: [](https://x.com/tavus/status/2105704984914559441)
- 2026-10-01, Hassaan Raza: [Tavus CEO Hassaan Raza on Griffin: 'I had no idea I was watching our model'](https://x.com/hassaanraza/status/2105709935074185546)
- 2026-09-30, Jonah Manzano: [](https://x.com/jonah_manzano/status/2105196063007334658)
- 2026-09-20, Qwen: [Qwen: Meet Qwen-Image-2.1, now open weights](https://x.com/Alibaba_Qwen/status/2101659302792679789)
- 2026-09-20, Chubby: ['An open-weight 7B model that outperforms Nano Banana 2.0'](https://x.com/kimmonismus/status/2101679965590769872)