Claude Opus 5.5 One-Shot a Music Video: What "Plan a Video" Actually Produces

A signal posted to X on September 23, 2026 reads: "Claude's Plan a video by opus 5.5 one shotted," with a tribute to @jeffgwoah attached. That is a demo claim, not a release claim — and it lands on the wrong side of the thing that usually matters. Claude Opus 5.5 is not an unreleased model to be sniffed out from a leak. Anthropic shipped it on September 22, 2026, at $4 per million input tokens and $20 per million output. It is a day old, generally available, and already on the rate card. So the interesting question is not whether Opus 5.5 exists. It is what these "one-shotted a video" posts are actually showing, because the phrase means something narrower than it sounds, and the difference decides whether you can use any of it.

Here is the short version, and it is the part most reposts leave out: in the demos that hold up to inspection, Claude Opus 5.5 does not generate pixels. It writes code that paints pixels. The output is a program, not a video file.

What the two demos in circulation actually did

Two public artifacts sit behind this class of claim, and both are checkable because both published their source.

The first is a 156.6-second music video for a song called "I'm Upping My P(doom)," a repo called PDoomVideo published on GitHub on September 22, 2026 — the same day Opus 5.5 launched. Its README states the video "took two generations, both in Claude Code," the first credited to Claude Opus 5.5 (Medium) and the second to Claude Opus 5.5 at default effort. It also states that "everything in this repository was generated by the model" and that "no scene ideas were specified" — the only direction given was to use a particular character design and to give each lyric interesting visuals and transitions.

The second is a three-game demo repo published September 23, 2026, which is closer to a controlled experiment than a showcase: three playable 3D browser games, one single-sentence prompt each, generated in a single session with what the repo describes as zero human edits to the code, then deployed. The games are single-file HTML with no external assets — models, textures, animation and sound are all generated procedurally by code. One of the three required the model to look up a game map's official layout and rebuild its geometry from that research before writing anything.

Neither repo is a vendor publication. Both are third-party artifacts, self-reported, and the "zero human edits" claim in particular is a claim about a git history, not something an outside party has audited. Treat the process description as the author's account and the artifacts themselves as the evidence.

The storyboard is the real output

The detail that decides whether "one-shotted" is a fair description is a file in the first repo called STORYBOARD.md. It is not a human's shot list. It is a document the model wrote for itself after its first generation pass, and it is genuinely specific: a rule that "something happens on screen" in every shot, an instruction to keep text out of frames, a named cast of characters with fixed designs, a palette that moves from warm cream through space violet to alarm red and back, a rule that character expressions morph rather than snap, and a requirement that every cut be motivated by action — a chomp to black, a fall, a zoom through an eye.

That document is the plan the signal text is pointing at. The model produced a directorial brief, then ran subagents against it in parallel, one per chapter, each writing a JavaScript file that paints its own segment of the timeline.

The rendering pipeline is ordinary and worth stating plainly, because it is where the "AI made a video" framing collapses: frames are painted in a page using p5.js and a watercolour brush library, a Node script drives that page in headless Chrome and screenshots every frame, and ffmpeg joins the frames with the audio track. At 24 frames per second across 156.6 seconds, that is roughly 3,760 frames rendered one at a time.

So the accurate sentence is not "Claude Opus 5.5 generated a video." It is "Claude Opus 5.5 wrote, and then directed, a program that renders a video." The distinction is not pedantic — it is the whole reason the technique is useful to anyone who is not making a music video.

Why the "two generations" line matters more than the "one-shot" line

The same README undercuts its own headline. The video took two generations, not one. The first pass produced a result the author then pushed further; the storyboard and the animation guide — the documents that make the second pass coherent — were written after that first pass, not before it.

That is the honest shape of every serious long-horizon generation task, and it is worth saying because the one-shot framing sets an expectation that the artifacts do not meet. The prompt was one message. The work was not one pass. What the model did was hold a large, multi-file, multi-hour build together well enough that the second pass was a revision rather than a restart — which is a real and much less flashy capability than "one prompt, one video."

There is an independent number that describes this better than the demos do. Artificial Analysis measured Claude Opus 5.5 on its Intelligence Index at 58 at maximum effort — the highest score it has recorded, several points clear of the next models — and reported that the model consumes roughly 119,000 output tokens per Index task, against about 73,000 for Claude Opus 5 and about 78,000 for Claude Fable 5.1. It uses around 1.6x the output tokens and lands at roughly the same cost per task (~$5.98 against ~$5.86) despite a 20% lower per-token price. Long-horizon generation is what that token profile looks like from the outside: the model is spending the budget on itself.

Where the artifacts stop being usable

The failure modes in this class of work are consistent, and both repos point at the same one from different directions.

• The output is a program, not a file. If you need an MP4 you can hand to someone, you still need Node, a browser, and ffmpeg. The model produced the renderer, not the render. That is a build dependency you own forever.

• It scales with frames, not with prompts. 3,760 frames took a scripted overnight pass on the author's own hardware. Nothing about Opus 5.5 changes the cost of painting frame 2,000.

• Determinism is a requirement, not a nicety. Frames render in parallel and out of order, so every frame must be a pure function of its timestamp — no carried state, no unseeded randomness. A model that has not internalised that produces a video that flickers. Both repos handle it; nothing in the prompt forces it.

• One-shot does not mean unaudited. The clearest evidence is the second repo's game work: the model had to research an existing map's layout before building it, which is the model deciding its first answer would have been wrong.

• Nobody has priced the failure. Neither repo reports how many attempts it took, how many tokens it burned, or how much of the second generation was fixing the first. That is the number a team actually needs before committing.

What this changes for you, and what it does not

If you were hoping for a video-generation model, Claude Opus 5.5 is not one, and no amount of demo footage makes it one. Anthropic's launch material contains no video-generation capability; the model's published evaluations are agentic coding, computer use and knowledge work. What the demos demonstrate is long-horizon code generation with a creative brief — the same faculty that shows up in the 680,000-line migration and the C-to-Rust port Anthropic cites in its own launch post, just pointed at something you can watch.

If that is the faculty you want, the practical question stops being "which model" and becomes "how do I run a thing like this without committing a production path to it." Long-horizon generation is expensive and its failure mode is silent — you get a plausible program that renders for four minutes and then breaks. The reasonable way to test it is to route the work through an endpoint you already have rather than a new contract: Opus 5.5 is on OrcaRouter at Anthropic's list price with 0% markup, alongside the other models in the family, so an overnight generation run can sit on the same key and the same failover rules as everything else you call. If the run dies at frame 3,000, that is a routing problem rather than a re-architecture.

Two things are worth watching before you plan around this. Anthropic has said Sonnet 5.5 and Haiku 5.5 follow "in the coming weeks," which will change the arithmetic on a long render — a cheaper model that can hold a multi-file build together would make the overnight pass routine rather than notable. And the storyboard pattern itself is portable: nothing in that document is specific to Opus 5.5, and nothing stops a smaller model from writing a worse one.

For now, the honest read of "Claude's Plan a video by opus 5.5 one shotted" is narrower than it sounds and more useful than it looks. The model wrote a shot list, briefed its own subagents, and then wrote the renderer — and the fact that a plan like that survived contact with 3,760 frames is the claim worth testing, not the video.