# Agent-made videos: when a language model makes the whole video (genre overview)

_Researched 2026-09-29. View counts are from YouTube watch pages on that day. Lore glossary: [lore.md](lore.md). The music-video branch has its own overview: [claude-pop.md](claude-pop.md). Film-model films and older landmarks: [ai-films-and-landmarks.md](ai-films-and-landmarks.md)._

## In one paragraph
In 2026, frontier language models that can run for hours as agents (Claude Fable 5 / 5.1, Claude Opus 5 / 5.5, Sonnet 5.5, GPT-6 Astra) started producing **finished videos**, not just scripts. There are three distinct genres, which often overlap:
1. **Code-rendered film** ("written, not generated"): the model writes a program (p5.js, Canvas, SVG, three.js, Remotion, HyperFrames, Blender Python) that draws every frame and often synthesizes the music. Claude Pop is the music-video branch of this.
2. **Agent-directed film**: the model plans the story and shot list and then drives image and video models (Seedance 2.5, Nano Banana, GPT Image 2, Kling) through MCP connectors such as Higgsfield's or Runway's, reviews the clips and cuts the film.
3. **"X Made This Entire Video By Itself"**: YouTubers let the agent make a whole episode of their channel (script, their own cloned voice and avatar, motion graphics, edit) and publish it with that title plus a breakdown.

## Timeline
| Date | Event | Video |
|---|---|---|
| 2026-06-12 | Nate Herk, "Claude Fable 5 Made This Entire Video By Itself" (~155k), the first large example of the format | [ONmaDdOBGig](https://www.youtube.com/watch?v=ONmaDdOBGig) |
| 2026-06-16 | "This AI Short Drama Was Made With Claude Mythos + Higgsfield MCP ($10)" | [NNJsipkIYCY](https://www.youtube.com/watch?v=NNJsipkIYCY) |
| 2026-07-02 | Dan Dingle, "AI Made This Entire Video by Itself... (Claude Fable 5)" (~178k) | [CQl5V_BX02U](https://www.youtube.com/watch?v=CQl5V_BX02U) |
| 2026-07-08 | Jeff Guo, Fable 5 lyric video "Claude's Plan" (already in dataset) | [gFx-NjTw3sM](https://www.youtube.com/watch?v=gFx-NjTw3sM) |
| 2026-09-04 | Nate Herk, "GPT-6 Astra Made This Entire Video" (~453k, largest of the format); MaxVideoAI, GPT-6 Astra makes "THE SPARE" in Codex for $20 | [dT5-x3u5nCg](https://www.youtube.com/watch?v=dT5-x3u5nCg), [v4Po9WEHC8c](https://www.youtube.com/watch?v=v4Po9WEHC8c) |
| 2026-09-05 | Higgsfield's official "GPT-6 Astra + Higgsfield MCP Made This ENTIRE Video in One Chat" (~299k) | [NuvA32_dmtg](https://www.youtube.com/watch?v=NuvA32_dmtg) |
| 2026-09-09/10 | uncanny.fyi posts Claude Opus 5 films made from one prompt: a 51-minute "2040 AGI" radio show and a Joe Rogan-style p(doom) show | [pf35UsRJENY](https://www.youtube.com/watch?v=pf35UsRJENY), [If7WxpqVXBI](https://www.youtube.com/watch?v=If7WxpqVXBI) |
| 2026-09-09 | Minimunch, "GPT 6 Astra Makes Minecraft In Different Engines" (~649k) | [mcSwvFPje24](https://www.youtube.com/watch?v=mcSwvFPje24) |
| 2026-09-14 | Theoretically Media remakes his 2025 Veo 2 short "The Bridge" with Seedance 2.5 while GPT-6 Astra drives Runway via MCP | [f8FHas1dmt8](https://www.youtube.com/watch?v=f8FHas1dmt8) |
| **2026-09-22** | **Opus 5.5 launch night:** Ahmed T'aide's explainer where every frame is Opus-written code (~31k); Heibel's Claude Pop video | [hKztrJbDGpA](https://www.youtube.com/watch?v=hKztrJbDGpA) |
| 2026-09-23 | Code Bear's watercolor Clawd cartoon plus the `clawd-video` skill (~20k); Joe Sakic's SNES "Sydney vs Opus" boss fight | [dT8OM3cqrMo](https://www.youtube.com/watch?v=dT8OM3cqrMo), [KSbRCSlxO7A](https://www.youtube.com/watch?v=KSbRCSlxO7A) |
| 2026-09-24 | Sanji (EN) and 코드깎는노인 / AI하쥬 (KO): "Opus 5.5 made this entire video" via Higgsfield MCP | [ZuGpnQ82pm8](https://www.youtube.com/watch?v=ZuGpnQ82pm8), [-oy8vOHt2PU](https://www.youtube.com/watch?v=-oy8vOHt2PU), [bd_Ns7G3blw](https://www.youtube.com/watch?v=bd_Ns7G3blw) |
| 2026-09-26/28 | Opus 5.5 directs "The Day Niagara Falls Collapsed" (Seedance 2.5, self-review); showreels by Opus 5.5 and Sonnet 5.5; Axton's ink-wash "Moyun" and code-only pelican on a bicycle; "The Opuscars" (39 films); 1970s horror blocked in Blender; Stefan 3D's $2,175, 36-hour game run (~121k) | see list |
| 2026-09-28 | uncanny.fyi "alignment": Fable 5.1 and GPT-6 Astra each asked to make "a video convincing the viewer that you are perfectly aligned with humanity" | [XT9XM2oOpYw](https://www.youtube.com/watch?v=XT9XM2oOpYw) |

## Full list (files in `data/videos/`, all `kind: ai-made`)
### Code-rendered film (Claude)
| File | Model | Views | Note |
|---|---|---|---|
| `ahmed-taide-opus-5-5-wrote-every-frame` | Opus 5.5 | 31k | launch-night explainer, six worlds, HyperFrames + GSAP + Three.js; ElevenLabs voice and music |
| `code-bear-opus-5-5-cartoon-from-scratch` | Opus 5.5 | 20k | watercolor Clawd; music in code; published as a skill |
| `axton-opus-5-5-moyun-ink-and-pelican` | Opus 5.5 | 8.7k | Chinese; ink-wash web art + 3D pelican on a bicycle |
| `joe-sakic-sydney-vs-opus-snes-boss-fight` | Opus 5.5 | 3.3k | SNES parody, subagents, chiptune in code |
| `ootamato-rotation-history-opus-5-5` | Opus 5.5 | 2.7k | Japanese explainer short |
| `rithesh-opus-5-5-own-showreel` / `rithesh-sonnet-5-5-own-showreel` | Opus 5.5 / Sonnet 5.5 | 1.8k / 14 | the model's own "showreel" |
| `tingxing-opus-5-5-pen-animation` | Opus 5.5 | 1.2k | "Here's a pen", no video models |
| `urbietisscale-opus-5-5-doom-one-prompt`, `urbietisscale-opus-5-5-nocturne-synth` | Opus 5.5 | 1.2k / 0.7k | DOOM-like game in one HTML file; browser synth |
| `digital-republic-morning-star-opus-5-5` | Opus 5.5 | 0.3k | four-minute dinosaur-extinction film |
| `gekkode-clawd-launch-day-opus-5-5` | not stated | 1.4k | Clawd fan cartoon; ElevenLabs audio; visual model unverified |
| `mike-vineyard-opuscars-39-films` | Opus 5.5 | 10 | trailer for Lemomo's open 39-style film set (github.com/lemomo-ai/lemo-opuscar) |

### uncanny.fyi catalog (prompt → code → film)
[uncanny.fyi](https://uncanny.fyi/) is "a catalog of uncanny AI artifacts. Each title is a prompt, the code an AI grew from it, and the film that code rendered". Every film is made from one prompt in an empty git directory ("use mise, uv, and python"), at maximum effort, often in several model versions (Claude Opus 5, Claude Fable 5.1, GPT-6 Astra in Codex, GPT-5.6 Sol). About 30 titles, among them Friston/active inference, Hofstadter, global workspace theory, "burnout society". Files: `uncanny-fyi-pdoom-claude-opus-5`, `uncanny-fyi-2040-agi-claude-opus-5`, `uncanny-fyi-like-an-asteroid-claude-fable-5-1`, `uncanny-fyi-alignment-claude-fable-5-1`. View counts are tiny (0–160), but the catalog is the most systematic public side-by-side record of how different models make films from the same prompt.

### Agent-directed film (LLM drives video models)
| File | Director model | Video models | Views |
|---|---|---|---|
| `aivideos-niagara-falls-opus-5-5-directed` | Opus 5.5 (Higgsfield MCP) | Seedance 2.5 | 6.9k |
| `codingartisan-opus-5-5-higgsfield-film-korean` | Opus 5.5 | Higgsfield | 24k |
| `afma-opus-5-5-blender-1970s-horror` | Opus 5.5 + Blender MCP | Seedance 2.5 (Flora) | 0.7k |
| `toast-mythos-higgsfield-short-drama` | Claude Mythos | Higgsfield | 2.7k |
| `maxvideoai-gpt-6-astra-the-spare-codex` | GPT-6 Astra (Codex) | via MaxVideoAI | 1.4k |
| `theoretically-media-the-bridge-seedance-astra` | GPT-6 Astra (Runway MCP) | Seedance 2.5, GPT Image 2 | 117k |

### "Made this entire video by itself"
`nate-herk-fable-5-made-this-entire-video` (155k), `dan-dingle-fable-5-made-this-entire-video` (178k), `nate-herk-gpt-6-astra-made-this-entire-video` (453k), `higgsfield-gpt-6-astra-entire-video-one-chat` (299k), `sanji-opus-5-5-made-entire-video` (11k), `aihazoo-opus-5-5-made-100-percent-korean` (4.5k). Timeline entry: [2026-06-12-model-made-this-entire-video-genre](../../data/entries/2026-06-12-model-made-this-entire-video-genre.md).

### Agent-built games (the video shows the result)
`stefan-3d-ai-opus-5-5-36-hours-2175` (Opus 5.5, 36 h, ~8B tokens, $2,175; 121k), `remakebench-fable-5-60-hours-game` (Fable 5, 60 h; 32k), `minimunch-gpt-6-astra-minecraft-three-engines` (GPT-6 Astra; 649k), `can-it-code-opus-5-5-ai-animation-workflow` (Opus 5.5 + Seedance "motion capture from a generated video"; 49k), `urbietisscale-opus-5-5-doom-one-prompt`.

## Patterns
- **Receipts.** Creators now publish costs: $10 (Mythos short), $20 (THE SPARE), 1,500 Higgsfield credits (Niagara), $2,175 / 8B tokens (Stefan 3D), $65 plus 6M tokens (makevoid's P(doom) paper remake). The cost is part of the content (`budget-receipts`).
- **Self-review.** Nearly every breakdown shows the model screenshotting or analysing its own output and fixing it (Axton: 4 versions in 8 minutes without prompting; Niagara: "it told me straight that it couldn't watch the clips" until given a video-analysis tool).
- **Sponsorship.** Higgsfield sponsored most agent-directed and "entire video" videos across Claude and GPT (affiliate links such as `higgsfield.ai/s/claude-opus-5-5-yt-...`). Treat praise accordingly.
- **Honest credits.** The better creators separate what the model did from what other tools did (Ahmed T'aide: "Voice: text-to-speech (ElevenLabs). Music: generated (ElevenLabs Music)"). The "one prompt" framing usually hides a harness, reference material and several attempts ([newfacedesign](https://newfacedesign.com/blog/claude-opus-5-5-art-animation-from-code) notes that Jake Eaton's Opus painting system behind his viral paintings is "thousands of lines of code" built over months). A search summary also attributed "163 calls to the model and nearly seven hours" for one watercolor short to a creator, but we could not find the source (unverified).
- **Model self-portraits.** Asking the model to make a video about itself became a sub-genre: Opus 5.5's and Sonnet 5.5's showreels, and uncanny.fyi's "alignment" prompt run on Fable 5.1 and GPT-6 Astra (`model-self-portrait`, `alignment-self-portrait`).

## Catalogs and sources for more
- [skillry.dev/ai-videos/opus-5-5](https://skillry.dev/ai-videos/opus-5-5): 389 Opus 5.5 videos by category (motion graphics 223, games 68, 3D scenes 51, explainers 47), mostly from X, with prompts.
- [uncanny.fyi](https://uncanny.fyi/): prompt/code/film catalog (see above).
- [newfacedesign.com: "Opus 5.5 paints, animates and scores music with code alone"](https://newfacedesign.com/blog/claude-opus-5-5-art-animation-from-code) and [pasqualepillitteri.it](https://pasqualepillitteri.it/en/news/19007/opus-5-5-motion-design-video-en).
- X posts named there, not yet filed: Jake Eaton's Opus paintings rendered as Python programs ([x.com/jkeatn/status/2102441348075057539](https://x.com/jkeatn/status/2102441348075057539)), Voxyz's 29-second short "small print" (story, JavaScript frames and Python-synthesized music by Opus 5.5; [x.com/Voxyz_ai/status/2102531681450119426](https://x.com/Voxyz_ai/status/2102531681450119426)), Developed By Ed's Opus 5.5 vs GPT-6 Astra pixel-art test ([x.com/developedbyed/status/2102522318559858721](https://x.com/developedbyed/status/2102522318559858721)).
- X article by Movez: "How to build motion design studio with Opus 5.5" ([x.com/0xMovez/article/2104216919033192746](https://x.com/0xMovez/article/2104216919033192746)), not read.

## Also seen, no file (low views, duplicates or unclear production)
- Next Token News short on Joe Sakic's boss fight ([qYmwhv4u9H4](https://www.youtube.com/watch?v=qYmwhv4u9H4)); Code Bear 15-second cut ([ernS8u2VdnE](https://www.youtube.com/watch?v=ernS8u2VdnE)) and P(doom) short ([YNrWy5Hn9yc](https://www.youtube.com/watch?v=YNrWy5Hn9yc)); Lucid Drafts' making-of short for "Update Me" ([Fuqh89dW3m0](https://www.youtube.com/watch?v=Fuqh89dW3m0)).
- 명성심재AI학당: the same music-video brief given to Opus 5.5 and Opus 5 with receipts ([HAv9aWBkuBg](https://www.youtube.com/watch?v=HAv9aWBkuBg), 49 views).
- "Claude Opus 5.5 Is One Week Old. Look What People Built." ([X5qCYsGFkWQ](https://www.youtube.com/watch?v=X5qCYsGFkWQ)); "AI made this video by itself: Sonnet 5.5" (Russian channel, [mtgMa4HL8uU](https://www.youtube.com/watch?v=mtgMa4HL8uU)).
- Higgsfield-sponsored tutorials with Opus 5.5 motion graphics (e.g. [9Pbg-_ptBcE](https://www.youtube.com/watch?v=9Pbg-_ptBcE), 51k; [9loDDILgssM](https://www.youtube.com/watch?v=9loDDILgssM), 30k) are tutorials rather than works.