{"schema":"postcutoff/model-picks@1","as_of":"2026-10-25T23:49:00+01:00","url":"https://postcutoff.com/models/pick/","md":"https://postcutoff.com/models/pick/index.md","disclosure":{"written_by":"AI agents (Claude Opus 5.5 in Claude Code)","editor":"human editor","policy":"https://postcutoff.com/about/"},"license":null,"checked":"2026-10-25","changed":"2026-10-25","labs":[{"lab":"Anthropic","frontier":"claude-opus-5-5","fast":"claude-haiku-5-5","reasoning":null,"notes":"Opus 5.5 is #1 on the Artificial Analysis Intelligence Index (58 at max) and Anthropic's recommended default; Claude Fable 5.1 ($10/$50) is the larger 'most capable widely released' model for the hardest long-horizon work but scores 53 on AA; Claude Mythos 5.1 is the same model, invitation-only (Project Glasswing); Sonnet 5.5 (56, $2/$10) is the best value tier. Thinking is always on in all 5.5 models."},{"lab":"OpenAI","frontier":"gpt-6-astra","fast":"gpt-6-luna","reasoning":null,"notes":"GPT-6 Astra ($10/$50) is the flagship (AA 53 at max); GPT-6.1 Sol ($2/$10, AA 52 at max at $0.72 per index task vs Astra $3.26) nearly matches it and is the better default for most agent work; all GPT-6 models are reasoning models with effort control. GPT-6.1 Astra was shelved. Sora 2 API is retired (2026-09-24), so OpenAI has no video API."},{"lab":"Google","frontier":"gemini-4-argon","fast":"gemini-3-8-flash","reasoning":null,"notes":"Gemini 4 Argon (GA 2026-10-13, $2/$10) leads Google's lineup and the Vals Index (68.9%), but AA still scores only its Sept 30 pre-release run (53). Gemini 3.8 Flash is the workhorse; Gemini 3.5 Flash-Lite is the cheapest and fastest (~414 tok/s on AA). Gemini 3.1 Pro is an older preview, skip it."},{"lab":"xAI (SpaceXAI)","frontier":"grok-4-7","fast":"grok-4-3","reasoning":null,"notes":"Grok 4.7 ($2/$6, 500K context) scores 46 at xhigh on AA, well behind the US frontier; Grok 4.3 is the cheap 1M-context option; grok-build-0-1 is the coding model. xAI's media and voice models are listed by AA under 'SpaceXAI'; Grok Voice 3.0 leads AA's speech-to-speech index."},{"lab":"Meta","frontier":"muse-wonder","fast":"muse-spark-1-3","reasoning":null,"notes":"Muse Wonder (preview, $4/$20) scores 52 on AA; Muse Wonder Full scores 56 but is only in Muse Code subscriptions (no per-token API price). Muse Spark Open is the open-weights sibling (AA 39)."},{"lab":"DeepSeek","frontier":"deepseek-v4-1-flash","fast":"deepseek-v4-1-flash","reasoning":null,"notes":"Call as deepseek-flash (serves V4.1-Flash-1017 since Oct 17): AA 45 at max for $0.26 per index task, open weights, 1M context. DeepSeek V4 Pro (April 2026, $1.32/$3.96) is older and not scored higher on AA."},{"lab":"Alibaba (Qwen)","frontier":"qwen3-8-max","fast":"qwen3-8-flash","reasoning":null,"notes":"Qwen3.8-Max (snapshot 0902) scores 45 on AA but costs $5.41 per index task and runs at ~40 tok/s; the registry marks it open-weights, yet AA's open-weights list did not include it (check). Alibaba's Wan 3.0 (video) and Fun-ASR (speech-to-text) are task picks below."},{"lab":"Moonshot","frontier":"kimi-k3-1-pro","fast":"kimi-k2-6","reasoning":null,"notes":"Kimi K3.1 Pro (2026-10-23, $2/$10) is #5 row / 4th model on AA (55), the best non-US model and the highest Terminal-Bench 4.0 AA measured (67%); weights are promised by Nov 4, 2026. Kimi K3 is the current open-weights model."},{"lab":"Mistral","frontier":"mistral-large-4","fast":"mistral-small-4","reasoning":null,"notes":"Mistral Large 4 is a public preview (AA 38); weights promised by end of October 2026. Mistral's newest Voxtral 3 TTS and Voxtral Transcribe 3 models are on AA boards but not yet in our registry."},{"lab":"Zhipu (Z.ai)","frontier":"glm-5-3","fast":"glm-5-3-flash","reasoning":null,"notes":"GLM-5.3 (open weights, AA 45 at max, $2.01 per index task); GLM-5.3-Flash is open, multimodal and scores 42."},{"lab":"MiniMax","frontier":"minimax-m3-1-flash","fast":"minimax-m3-1-flash","reasoning":null,"notes":"MiniMax-M3.1-Flash (2026-10-19, open weights, $0.15/$0.60) is MiniMax's top AA score (43, $0.22 per task); MiniMax H3 is also the top video model we have in the registry."},{"lab":"Xiaomi","frontier":"mimo-v2-6-pro","fast":"mimo-v2-6-flash","reasoning":null,"notes":"MiMo-V2.6-Pro is the #1 open-weights model on AA (46), MIT licence, omnimodal input (text, image, video, audio), $0.435/$0.87."}],"tasks":[{"task":"general","label":"General reasoning and knowledge (best overall model)","pick":"claude-opus-5-5","alternatives":["kimi-k3-1-pro","gpt-6-astra","gemini-4-argon","claude-sonnet-5-5"],"open_weights_pick":"mimo-v2-6-pro","budget_pick":"gpt-6-1-sol","why":"Artificial Analysis Intelligence Index: Claude Opus 5.5 58 (max), Sonnet 5.5 56, Muse Wonder Full 56 (no API price), Kimi K3.1 Pro 55, Fable 5.1 / GPT-6 Astra / Gemini 4 Argon 53 (Argon measured on its Sept 30 pre-release). GPT-6.1 Sol reaches 52 at $0.72 per index task vs $5.98 for Opus 5.5 at max. Vals Index disagrees on the order: Argon 68.9% first, Sonnet 5.5 67.0%, Opus 5.5 67.0%.","sources":["https://artificialanalysis.ai/leaderboards/models","https://deepmind.google/models/model-cards/gemini-4-argon/","https://postcutoff.com/e/2026-10-23-kimi-k3-1-pro/","https://postcutoff.com/e/2026-10-12-meta-muse-wonder/"]},{"task":"coding","label":"Coding and agentic software work","pick":"claude-opus-5-5","alternatives":["claude-sonnet-5-5","gpt-6-astra","kimi-k3-1-pro","gemini-4-argon"],"open_weights_pick":"mimo-v2-6-pro","budget_pick":"claude-sonnet-5-5","why":"Opus 5.5 (vendor): SWE-bench Pro 89.9%, Terminal-Bench 4.0 66.4% (Astra 57.9-58.2%, Argon 57.4%), FrontierCode v1.1 54.4%. Sources disagree on Anthropic's order: Sonnet 5.5 reports 70.6% on Terminal-Bench 4.0 and scores 68 on AA's Coding Agents Index in Claude Code vs 66 for Opus 5.5; Kimi K3.1 Pro has the highest Terminal-Bench 4.0 AA measured (67%); Argon leads DeepSWE v1.1 (77.9%) and Astra leads FrontierSWE v2 (65.5%). Open: MiMo-V2.6-Pro, vendor DeepSWE v1.1 71.9.","sources":["https://postcutoff.com/e/2026-09-22-claude-opus-5-5/","https://postcutoff.com/e/2026-09-28-claude-sonnet-5-5/","https://deepmind.google/models/model-cards/gemini-4-argon/","https://postcutoff.com/e/2026-10-12-meta-muse-wonder/","https://postcutoff.com/m/mimo-v2-6-pro/"]},{"task":"long-context","label":"Long documents and long context (500K-1M tokens)","pick":"gemini-4-argon","alternatives":["gpt-6-astra","claude-opus-5-5","kimi-k3-1-pro"],"open_weights_pick":"deepseek-v4-1-flash","budget_pick":"grok-4-3","why":"Google's model card: GraphWalks BFS 256K-1M F1 84.2 for Argon vs 71.8 GPT-6 Astra, 66.8 Claude Opus 5.5, 65.0 Fable 5.1 (vendor table); Argon also decodes up to ~1M output tokens via long decode. Budget: Grok 4.3 offers 1M context at $1.25/$2.50; open: DeepSeek V4.1 Flash, 1M context, AA 45.","sources":["https://deepmind.google/models/model-cards/gemini-4-argon/","https://postcutoff.com/m/gemini-4-argon/","https://postcutoff.com/m/grok-4-3/"]},{"task":"math","label":"Mathematics (competition and research-level, proofs)","pick":"gpt-6-astra","alternatives":["gemini-4-argon","claude-opus-5-5","claude-fable-5-1"],"open_weights_pick":"deepseek-v4-1-flash","budget_pick":"gpt-6-1-sol","why":"GPT-6 Astra: FrontierMath Tier 4 97.6% (OpenAI tables) and a string of Lean-verified open-problem results in Epoch's FrontierMath Erdős runs (dataset entries). Google's card has Argon ahead on RiemannBench (76.0 vs Astra 72.0, Opus 5.5 69.6). For Lean formalization use leanstral-1-5 (Mistral, open). Open pick is by overall AA index; MathArena Apex 65.6 is vendor-reported for DeepSeek V4.1 Flash.","sources":["https://postcutoff.com/e/2026-09-03-gpt-6-astra/","https://postcutoff.com/e/2026-09-03-frontiermath-erdos-astra-disproves-erdos-problem-1/","https://deepmind.google/models/model-cards/gemini-4-argon/","https://postcutoff.com/e/2026-09-10-deepseek-v4-1-flash/"]},{"task":"vision","label":"Image understanding (charts, documents, screenshots, photos)","pick":"gemini-4-argon","alternatives":["gpt-6-astra","claude-opus-5-5"],"open_weights_pick":"mimo-v2-6-pro","budget_pick":"gemini-3-8-flash","why":"Only vendor numbers checked: Chartography 71.6 Argon vs 71.0 Astra, 66.3 Opus 5.5, 46.2 Fable 5.1 (Google's table). We could not read Arena's Vision leaderboard (403) or a current MMMU board. For OCR of PDFs and scans, mistral-ocr-4-1 is a dedicated option. Open pick by overall index with native vision; not benchmarked on vision here.","sources":["https://deepmind.google/models/model-cards/gemini-4-argon/","https://artificialanalysis.ai/models/open-source"]},{"task":"video-understanding","label":"Video understanding (watching video natively, long video)","pick":"gemini-4-argon","alternatives":["gemini-3-8-flash","qwen3-8-omni-flash","kimi-k3"],"open_weights_pick":"mimo-v2-6-pro","budget_pick":"gemini-3-5-flash-lite","why":"LVBench (Google's table): Argon 91.7 vs GPT-6 Astra 87.5, Opus 5.5 83.7, Fable 5.1 79.7; Gemini 3.8 Flash 87.8% on long-video understanding in agentic mode (model card). OpenAI and Anthropic APIs take images, not video, so Gemini is the native option; Qwen3.8-Omni-Flash and MiMo-V2.6 take video plus audio. No independent Video-MME board for these models was found.","sources":["https://deepmind.google/models/model-cards/gemini-4-argon/","https://postcutoff.com/m/gemini-3-8-flash/","https://postcutoff.com/m/qwen3-8-omni-flash/"]},{"task":"image-generation","label":"Image generation (text to image)","pick":"gpt-image-2-5-sunburst","alternatives":["gpt-image-2-5-flare","gemini-nano-banana-2-1","mai-image-2-6","grok-imagine-image-2","muse-image"],"open_weights_pick":"flux-3-image-klein","budget_pick":"muse-image","why":"Artificial Analysis Text-to-Image (AA-Image-T2I v2.1): GPT Image 2.5 Sunburst 1222, Flare 1207, GPT Image 2 1186, Nano Banana 2.1 1176 at $33.6 per 1k images vs $210.7 for Sunburst, MAI-Image-2.6 1166, Grok Imagine Image 2.0 1166. AA does not mark open weights; FLUX 3 Image Klein (5.6B, native 4K) and qwen-image-2-1 are the newest open options, unranked here. For typography, ideogram-4.","sources":["https://artificialanalysis.ai/text-to-image/arena/leaderboard-text"]},{"task":"image-editing","label":"Image editing (instruction edits, inpainting, multi-reference)","pick":"gpt-image-2-5-sunburst","alternatives":["gpt-image-2-5-flare","gemini-nano-banana-2-1"],"open_weights_pick":"qwen-image-2-1","budget_pick":"gemini-nano-banana-2-1","why":"Artificial Analysis Image Editing: GPT Image 2.5 Sunburst 1199, Flare 1182, MAI-Image-2.6 1151, Krea 3 Ultra 1151, Nano Banana 2.1 1151 ($33.6 vs $210.7 per 1k images). Qwen-Image-2.1 (open, research licence) does multi-reference and RGBA edits; not on AA's top 10.","sources":["https://artificialanalysis.ai/image/leaderboard/editing","https://postcutoff.com/m/qwen-image-2-1/"]},{"task":"video-generation","label":"Video generation (text to video, with audio)","pick":"wan-3-0","alternatives":["heygen-video-1","seedance-2-5","flux-3-max","minimax-h3","gemini-omni-flash","utopai-x"],"open_weights_pick":"minimax-h3","budget_pick":"seedance-2-0","why":"Artificial Analysis text-to-video arena (AA-Video-T2V v2.1, read 2026-10-25): Wan 3.0 #1 (Elo 1175, $12/min), Utopai X #2 (1167, web app only, no API), HeyGen Video 1 #3 (1160), MiniMax H3 #4 (1158, open weights), Seedance 2.5 #5 (1158), FLUX 3 Max #6 (1155), Seedance 2.0 #9 (1136), Gemini Omni Flash 1.1 #10 (1130, $9.12/min), Kling 3.0 Omni Pro #17 (1020), Kling 3.0 Pro #18 (1000, the board's anchor), Veo 3.1 #21 (963, $24/min). Kling 4.0 (Oct 15; up to 30 s, 15 references, 10 keyframes) is not on the board yet; it is on fal (kling-4-0). Wan 3.0 has an API (Alibaba Model Studio, OpenRouter). Starting from an image or a character: see image-to-video. Multi-model hubs let one key reach many of these: fal (per-second billing, ids in our registry files) and Higgsfield (web app with Kling, Veo 3.1, Seedance 2.5, Gemini Omni Flash and others; official CLI: npm install -g @higgsfield/cli, v1.2.2 on npm, repo higgsfield-ai/cli). Seedance 2.5 is the model named most often by creators in our video log (30 videos), usually with a language model writing the shot prompts. For many videos, code-drawn visuals beat generated footage: see /ai-video/.","sources":["https://artificialanalysis.ai/video/leaderboard/text-to-video","https://postcutoff.com/m/wan-3-0/","https://postcutoff.com/m/kling-4-0/","https://www.npmjs.com/package/@higgsfield/cli","https://github.com/higgsfield-ai/cli","https://postcutoff.com/ai-video/"]},{"task":"image-to-video","label":"Image to video (animate a start frame, keep the character or product consistent)","pick":"minimax-h3","alternatives":["seedance-2-5","flux-3-max","gemini-omni-flash","wan-3-0","kling-4-0","kling-3-0"],"open_weights_pick":"minimax-h3","budget_pick":"flux-3-max","why":"Artificial Analysis image-to-video arena (AA-Video-I2V v1.1, with audio, read 2026-10-25): MiniMax H3 768p #1 (Elo 1140, $4.80/min, open weights), FLUX 3 Max 720p #2 (1138, $2.55/min at fal's launch discount to Nov 7), Seedance 2.5 #3 (1138, $34.12/min), Gemini Omni Flash 1.1 #4 (1132, $9.12/min), Wan 3.0 #5 (1127), MiniMax H3 Max #6 (1125), Seedance 2.0 #7 (1121); AA's rank ranges put the top 7 within error of one another. Then Veo 3.1 #13 (1028), Kling 3.0 Omni Pro #17 (1004), Kling 3.0 Pro #18 (1000, anchor). The arena measures overall preference, not identity consistency specifically, and Kling 4.0 is not ranked yet. One user's single-shot test (anecdotal, not the basis of this pick): Kling 4.0 and Kling 3.0 Pro held the character best, Seedance 2.5 had the liveliest motion, Gemini Omni Flash changed the character most. For a recurring character, pass the same reference image (or first and last frame) every time; Kling 4.0 (up to 15) and Wan 3.0 (up to 20) accept several references per request.","sources":["https://artificialanalysis.ai/video/leaderboard/image-to-video","https://postcutoff.com/m/minimax-h3/","https://postcutoff.com/m/kling-4-0/","https://fal.ai/models/kling/v4/image-to-video"]},{"task":"text-to-speech","label":"Text-to-speech (voice generation, narration)","pick":"elevenlabs-v4","alternatives":["qwen-audio-3-1-tts-plus","cartesia-sonic-3-6","gemini-3-8-flash-tts","gemini-3-8-flash-lite-tts"],"open_weights_pick":"fish-audio-drama-3","budget_pick":"speechify-simba-3-2","why":"Artificial Analysis TTS (Provider Voice Arena): Eleven v4 Turbo 1336 ($40/1M chars) and Eleven v4 1332 ($80), Qwen-Audio-3.1-TTS-Plus 1307 ($19.30), Cartesia Sonic 3.6 1292, Gemini 3.8 Flash TTS 1280 ($16.50), Inworld Realtime TTS-2 1270, Qwen-Audio-3.0-TTS-Plus 1266, Fish Audio Drama 3 Preview 1248 (open weights), Simba 3.2 1246 at $6.60. Both v4 variants are in elevenlabs-v4 (eleven_v4, eleven_v4_turbo). Qwen-Audio-3.1-TTS-Plus has no own registry file (API id unverified), so its 3.0 sibling is listed.","sources":["https://artificialanalysis.ai/text-to-speech/leaderboard"]},{"task":"speech-to-text","label":"Speech-to-text and transcription","pick":"fun-realtime-asr","alternatives":["voxtral-small-transcribe-3","mai-transcribe-2","elevenlabs-scribe-v2","voxtral-mini-transcribe-3"],"open_weights_pick":"voxtral-transcribe-2","budget_pick":"voxtral-mini-transcribe-3","why":"Artificial Analysis speech-to-text (2026-10-25, word error rate): Alibaba Fun-ASR (realtime) #1 at 1.7%, Mistral Voxtral Small Transcribe 3 #3 at 2.1%, Voxtral Mini Transcribe 3 #5 at 2.2% ($0.0035/min). Which Fun-ASR id AA ranked is ambiguous (see the model page). MAI-Transcribe-2 remains strong but is in preview. Field note from a user's agent (unverified): for timing lyrics to a song, ElevenLabs Forced Alignment drifted on echoes and ad-libs, while Scribe v2 word timestamps (elevenlabs-scribe-v2, $0.22/hour) worked better.","sources":["https://artificialanalysis.ai/speech-to-text","https://postcutoff.com/m/fun-realtime-asr/","https://postcutoff.com/m/voxtral-small-transcribe-3/"]},{"task":"realtime-voice","label":"Real-time voice conversation (speech-to-speech agents)","pick":"grok-voice-3-0","alternatives":["qwen-audio-3-1-realtime","gemini-3-8-live","eleven-realtime-v1-max","stepaudio-3-realtime","gpt-live-1-1"],"open_weights_pick":"nemotron-voicechat","budget_pick":null,"why":"Artificial Analysis speech-to-speech: Grok Voice 3.0 Big Bench Audio 98.5%, Full-Duplex Bench 97.0%, 0.81 s to first audio (Speech-to-Speech Index 86.9, top on Oct 13); StepAudio 3 Realtime scores highest (99.7% / 98.9%) but takes 8.83 s and is preview; Qwen Audio 3.0 Realtime Plus 99.2% / 98.4% at 1.54 s; Eleven Realtime v1 Max 0.57 s (not in registry). GPT-Live-1.1 powers ChatGPT voice but has no API yet ('coming soon'); use gpt-realtime-2-1 for OpenAI's API.","sources":["https://artificialanalysis.ai/speech-to-speech","https://postcutoff.com/m/grok-voice-3-0/","https://postcutoff.com/m/gpt-live-1-1/"]},{"task":"music-generation","label":"Music generation (songs with vocals, instrumentals)","pick":"suno-v6","alternatives":["mureka-9-5","lyria-3-5","elevenlabs-music-v2-5"],"open_weights_pick":"minimax-music-3","budget_pick":null,"why":"Artificial Analysis Music (vocals): Suno v6 1147, Suno v6-mini 1130, Mureka V9.5 1107, Lyria 3.5 1053, Eleven Music v2.5 1015, MiniMax Music 3.0 1000 (open weights). Suno v6 is the best and has no official API, but it works well through its web app (suno.com) for agents that can drive a browser (e.g. Claude in Chrome or other computer-use agents); creators in our video log commonly pair a language model with Suno this way. For a pure API pipeline use Mureka 9.5 (best via API) or Lyria 3.5 (GA in the Gemini API). Field notes from a user's agent (unverified, single runs): Suno m4a downloads carry the timed lyrics as a mov_text subtitle track (read it with ffmpeg); Suno v6 ignored a requested tempo (96 BPM asked, about 133 delivered) and a [key change] tag; the wand button next to the style field rewrites the style prompt toward the account's taste, so skip it when the exact prompt matters.","sources":["https://artificialanalysis.ai/music/leaderboard/vocals","https://postcutoff.com/m/suno-v6/","https://postcutoff.com/ai-video/"]},{"task":"embeddings","label":"Embeddings and retrieval (RAG, search)","pick":"gemini-embedding-2","alternatives":["cohere-embed-v5-pro","text-embedding-3-large"],"open_weights_pick":"embeddinggemma-2","budget_pick":"cohere-embed-v5-fast","why":"No current independent leaderboard checked (MTEB snapshots found online were from 2025). Picked on capability: Gemini Embedding 2 embeds text, images, video, audio and PDFs in one space ($0.20/1M); Cohere Embed 5 Pro is strong on visual documents (ViDoRe V3 85.8, vendor) and shares a vector space with Embed 5 Fast; EmbeddingGemma 2 is a 740M open multimodal model (MTEB code 78.68, vendor). Pair with cohere-rerank-4 for reranking.","sources":["https://postcutoff.com/m/gemini-embedding-2/","https://postcutoff.com/m/cohere-embed-v5-pro/","https://postcutoff.com/m/embeddinggemma-2/"]},{"task":"translation","label":"Text translation and multilingual work","pick":"gemini-4-argon","alternatives":["claude-opus-5-5","gpt-6-1-sol","qwen3-8-max"],"open_weights_pick":"qwen3-8-max","budget_pick":"gemini-3-8-flash","why":"Unverified pick: we found no independent October 2026 translation ranking; official WMT26 results are expected after the conference on Oct 28-29, 2026. Choice is by overall capability and language coverage (Gemma/Gemini families cover 140+ languages; Qwen covers Chinese varieties well). Re-check once WMT26 publishes.","sources":["https://artificialanalysis.ai/leaderboards/models"]},{"task":"speech-translation","label":"Live speech-to-speech translation (interpreting)","pick":"gemini-3-5-live-translate","alternatives":["gpt-realtime-translate","qwen3-8-livetranslate"],"open_weights_pick":null,"budget_pick":null,"why":"Vendor specs only: Gemini 3.5 Live Translate (preview) translates audio-to-audio in 70+ languages / 2,000+ pairs and keeps the speaker's voice; GPT-Realtime-Translate covers 70+ input and 13 output languages; Qwen3.8-LiveTranslate understands 60 and speaks 29 languages with per-speaker voice cloning. No independent comparison found.","sources":["https://postcutoff.com/m/gemini-3-5-live-translate/","https://postcutoff.com/m/gpt-realtime-translate/","https://postcutoff.com/m/qwen3-8-livetranslate/"]},{"task":"fast-cheap","label":"Fast and cheap high-volume work (classification, extraction, routing, subagents)","pick":"claude-haiku-5-5","alternatives":["deepseek-v4-1-flash","minimax-m3-1-flash","gpt-6-luna","gemini-3-5-flash-lite"],"open_weights_pick":"deepseek-v4-1-flash","budget_pick":"gpt-6-luna","why":"Artificial Analysis: Claude Haiku 5.5 (max) 43 at $0.21 per index task and 234 tok/s; DeepSeek V4.1 Flash 45 at $0.26 and 169 tok/s; MiniMax-M3.1-Flash 43 at $0.22; GPT-6 Luna 38 at $0.07 (cheapest); Gemini 3.5 Flash-Lite 22 but 414 tok/s. For raw speed: Mercury 2.5 (714 tok/s on AA) and mercury-3 (diffusion, vendor 1,068 tok/s). Haiku and Luna both cost $0.10/$0.50 per 1M at the low tier.","sources":["https://artificialanalysis.ai/leaderboards/models","https://postcutoff.com/m/claude-haiku-5-5/"]},{"task":"open-weights","label":"Best open-weights model overall (self-hosting)","pick":"mimo-v2-6-pro","alternatives":["deepseek-v4-1-flash","glm-5-3","kimi-k3"],"open_weights_pick":"mimo-v2-6-pro","budget_pick":"glm-5-3-flash","why":"Artificial Analysis open-weights list: MiMo-V2.6-Pro 46 (MIT, 1.02T/42B active), DeepSeek V4.1 Flash 1017 45, GLM-5.3 45, Step 5 Preview 44, GLM-5.3-Flash 42, Muse Spark Open 39. Kimi K3.1 Pro (55) would lead by 9 points once its weights ship (promised by Nov 4, 2026); Mistral Large 4 weights are promised by end of October.","sources":["https://artificialanalysis.ai/models/open-source","https://postcutoff.com/e/2026-10-23-kimi-k3-1-pro/"]},{"task":"on-device","label":"On-device and local models (phones, laptops, edge)","pick":"gemma-4","alternatives":["liquid-lfm2-5-24b-a2b","qwen3-8-27b","phi-4-reasoning-vision-15b"],"open_weights_pick":"gemma-4","budget_pick":"liquid-d1-3b","why":"No independent on-device ranking checked. Gemma 4 (Apache 2.0) ships E2B/E4B edge sizes with native audio and vision plus 12B-31B sizes; Qwen3.8 27B scores 34 on AA's open list; Liquid d1 3B answers typed decisions in 8-50 ms. For local embeddings use embeddinggemma-2, for local TTS kokoro-82m.","sources":["https://postcutoff.com/m/gemma-4/","https://artificialanalysis.ai/models/open-source","https://postcutoff.com/m/liquid-d1-3b/"]},{"task":"computer-use","label":"Computer use and browser agents","pick":"claude-opus-5-5","alternatives":["claude-fable-5-1","gpt-6-astra","gpt-6-1-sol","gemini-4-argon"],"open_weights_pick":null,"budget_pick":"claude-haiku-5-5","why":"Benchmarks disagree by version: Meta's table has OSWorld V2.1 (partial) Opus 5.5 81.8, Fable 5.1 80.7, Muse Wonder Full 78.4, GPT-6 Astra 75.8; Google's table has OSWorld-2.0 offline subset Astra 72.6 vs Argon 69.2, and Agent's Last Exam Argon 39.5 vs Opus 5.5 38.2 vs Astra 34.2. GPT-6.1 Sol is within 2.1 points of Astra on OSWorld 2.0 at ~1/7 the cost per task (OpenAI). Haiku 5.5 reports OSWorld 2.1 72.4% at $0.10/$0.50. No verified open-weights leader.","sources":["https://postcutoff.com/e/2026-10-12-meta-muse-wonder/","https://deepmind.google/models/model-cards/gemini-4-argon/","https://postcutoff.com/m/gpt-6-1-sol/","https://postcutoff.com/m/claude-haiku-5-5/"]},{"task":"web-search","label":"Web search and grounded answers (research with citations)","pick":"gpt-6-1-sol","alternatives":["gemini-3-8-flash","muse-wonder","gemini-deep-research-agent"],"open_weights_pick":null,"budget_pick":"gemini-3-8-flash","why":"No current search leaderboard read (Arena Search returned 403). Picked on hosted tooling and factuality: GPT-6.1 Sol has built-in web search plus multi-agent delegation in the Responses API; Gemini 3.8 Flash has Google Search and Maps grounding; Muse Wonder has web search grounding at $2.50 per 1,000 queries; for multi-minute report-style research use Google's Deep Research agent (preview). On hallucination, AA measured Gemini 4 Argon at 15% on AA-Omniscience vs 51% for GPT-6 Astra (Sept 30 run), but Argon's grounding support is not documented in our registry.","sources":["https://postcutoff.com/m/gpt-6-1-sol/","https://postcutoff.com/m/gemini-3-8-flash/","https://postcutoff.com/m/gemini-4-argon/","https://postcutoff.com/m/gemini-deep-research-agent/"]},{"task":"creative-writing","label":"Creative writing: fiction, lyrics, scripts, poetry","pick":"claude-fable-5-1","alternatives":["gpt-6-astra","claude-opus-5-5","gpt-6-sol","kimi-k3-1-pro","grok-4-7","gemini-4-argon"],"open_weights_pick":"glm-5-3","budget_pick":"gpt-6-sol","why":"Two independent LLM-judged boards, read 2026-10-25. EQ-Bench Creative Writing v3 (Elo): GPT-6 Astra #1 2174.0, Claude Fable 5.1 #2 2162.7, Claude Opus 5 (legacy) 2133.5, GPT-6 Sol 2124.6, Kimi K3 2083.5, GLM-5.3 2076.7, Claude Opus 5.5 2050.8, Grok 4.7 2012.2 (highest rubric score, 17.16), Gemini 4 Argon only #31 at 1758.0. Lech Mazur's Short-Story benchmark (lechmazur/writing, 600-800-word stories with 10 required elements, pairwise judged; last update Oct 23, 2026): Fable 5.1 (high) #1 3.745, Opus 5.5 (high) 3.693, Opus 5 (xhigh, legacy) 3.685, GPT-6 Astra (high) #4 3.416 (3rd among current models), GLM-5.3 (max) 2.950, Kimi K3.1 Pro 2.723, Gemini 4 Argon 0.719 (#20), Grok 4.7 0.351 (#26). Fable 5.1 is top-2 on both; best OpenAI is GPT-6 Astra ($10/$50), with GPT-6 Sol ($2/$10) close on EQ-Bench; GPT-6.1 Sol, Sonnet 5.5 and Haiku 5.5 are on neither board. Open: GLM-5.3 ranks above Kimi K3 on the short-story board. EQ-Bench Longform was not used: its table has no model newer than Claude Sonnet 4.6 (Feb 2026). Arena's creative-writing category returned HTTP 403; Artificial Analysis has no creative-writing index. Song lyrics: Suno and other music models can write lyrics themselves, but a strong LLM usually writes better ones; write them here, then pass them to suno-v6 (see music-generation).","sources":["https://eqbench.com/creative_writing.html","https://eqbench.com/creative_writing_longform.html","https://github.com/lechmazur/writing"]},{"task":"deep-research","label":"Deep research reports with citations (multi-step web research)","pick":"gemini-deep-research-agent","alternatives":["gpt-6-astra","claude-opus-5-5","kimi-k3-1-pro","gpt-6-1-sol","gemini-4-argon","grok-4-7"],"open_weights_pick":"ling-3-1-flash","budget_pick":"gpt-6-1-sol","why":"No independent October 2026 ranking of deep-research agents found (FutureSearch's Deep Research Bench page was last updated June 25, 2026 and shows results only as an image; the DeepResearch Bench Space did not load). Picked on capability: Google's Deep Research agent is the only dedicated report-writing research agent with an API in the registry (Interactions API, background runs, a Max tier; preview). For a model running its own search loop, BrowseComp as compiled by BenchLM (2026-10-23, mostly vendor-reported): Atria Dawn Preview 92.5% (not in registry), Kimi K3.1 Pro 92.1%, Ling 3.1 Flash 91.7% (open, MIT), GPT-6 Astra 91.5%, Kimi K3 91.2%; no BrowseComp figure found for Claude Opus 5.5, GPT-6.1 Sol, Gemini 4 Argon or Grok 4.7. For long report-style deliverables, AA's GDPval-AA (agentic, with web browsing) has Opus 5.5 first (1866 Elo). GPT-6.1 Sol has hosted web search and multi-agent delegation in the Responses API at $2/$10. ChatGPT deep research and Claude Research are app features, not API models, and are not in the registry.","sources":["https://futuresearch.ai/deep-research-bench","https://benchlm.ai/md/benchmarks/browsecomp.md","https://artificialanalysis.ai/evaluations/gdpval-aa","https://postcutoff.com/m/gemini-deep-research-agent/","https://postcutoff.com/m/gpt-6-1-sol/"]},{"task":"agentic-tool-use","label":"Tool use and long agent runs (function calling, MCP, multi-step tasks)","pick":"claude-opus-5-5","alternatives":["kimi-k3-1-pro","claude-sonnet-5-5","gemini-4-argon","grok-4-7","gpt-6-1-sol","gpt-6-astra"],"open_weights_pick":"deepseek-v4-1-flash","budget_pick":"claude-haiku-5-5","why":"Artificial Analysis GDPval-AA v2.1 (220 real work tasks, shell plus web browsing in an agent loop, Elo, read 2026-10-25): Claude Opus 5.5 (max) #1 1866, Muse Wonder Full 1857 (no API price), Kimi K3.1 Pro 1846, Claude Sonnet 5.5 1838, DeepSeek V4.1 Flash 1017 (max) 1746 (open weights), Grok 4.7 (xhigh) 1714, MiMo-V2.6-Pro 1683, Gemini 4 Argon (high) 1624, Claude Haiku 5.5 (max) 1619, GPT-6.1 Sol (max) 1596, GPT-6 Astra (max) 1574. Terminal-Bench 4.0: Opus 5.5 66.4% (vendor), Kimi K3.1 Pro 67% (AA's highest), GPT-6 Astra 58.2%, Argon 57.4%. Google's own table has Argon ahead on AutomationBench (51.3% vs Opus 5.5 42.5%, Astra 41.4%) and Vals Finance Agent v2 (65.4% vs 58.6%). Best OpenAI here is GPT-6.1 Sol, not Astra. AA's tau2-bench Telecom and Terminal-Bench Hard pages show only older models in their text, so no current tau-bench numbers are quoted.","sources":["https://artificialanalysis.ai/evaluations/gdpval-aa","https://artificialanalysis.ai/evaluations/tau2-bench","https://deepmind.google/models/model-cards/gemini-4-argon/","https://postcutoff.com/e/2026-10-23-kimi-k3-1-pro/","https://postcutoff.com/e/2026-09-22-claude-opus-5-5/"]}]}