Claude Next Coded a 51-Second Video: What the "2.5 Hours" Claim Actually Shows

On October 12, 2026, the X account @kimmonismus relayed a claim about an unreleased Claude model that does not exist on any vendor surface: that an upcoming, unnamed Claude model — the post's source calls it Claude Next — chose the images, the music and the visual direction for a video, then "coded the rest of the production in 2.5 hours," doing its own research and making its own creative decisions. The video itself was posted a day earlier by @chetaslua, captioned "Claude Next Model," with the claim that what you are watching is "fully coded" and that the model makes Claude Opus 5.5 "look primitive."

Start with what is actually verifiable in that, because it is more than the reposts suggest and less than the framing implies. The video is real, it is 51 seconds long, and it exists as a video file on X. Claude Next is not a model Anthropic has released, documented or priced — there is no model page, no rate card, no row in the models overview, and no vendor statement. And the one half of the claim that carries all the technical weight, that the "production was built with code" while the images and music were "existing assets," is the same shape of demo that Claude Opus 5.5 has already been publicly shown doing, at length, including by the same X account that posted this one four weeks earlier.

That last point is the reason this is worth a piece rather than a shrug. Anthropic's most advanced shipped model has spent September and October being demonstrated on exactly this: writing code that paints pixels rather than generating pixels directly. So "an unreleased model coded a video" is not a claim anyone has to take on faith about a model nobody can call. It is a claim about degree, and the degree is measurable — in frames, in tokens, in hours, and in the two things the video lets you inspect for yourself.

What was posted, and by whom

Two accounts, two posts, 16 hours apart, and neither of them published by Anthropic. Getting the chain right matters, because the two posts claim different things.

• The first, from @chetaslua at 14:32 UTC on October 11, shows a video attachment of 51.4 seconds at 1920×1080 and says: "Claude Next is being tested on claude code and i thought it will take a month to make image and video gen model obsolete i was wrong. Behold the supremacy, fully coded, this akes even Opus 5.5 look primitive. I will upload more test different types its SOTA in every f#ing thing." The post has roughly 438,000 views at the time of writing.

• The second, from @kimmonismus at 12:17 UTC on October 12, is a quote-post that paraphrases rather than reports: "Anthropics upcoming model is another insane leap. Claude reportedly chose the images, music and visual direction for this video, then coded the rest of the production in 2.5 hours. According to @chetaslua, it was free to do its own research and make the creative decisions." It then adds its own clarifying line, which is the most useful sentence in the whole exchange: "The images and music are existing assets. The surrounding production was built with code." That post has roughly 151,000 views; a search of the surrounding record found no additional test posts, no screenshots of a model picker, and no artifact beyond the video.

So the "2.5 hours" figure, the "chose the images, music and visual direction" framing, and the "free to do its own research" claim are all paraphrase at one remove from a person who says he tested the model and intends to publish more tests. There is no second independent tester in the record. Treat every number above as reported and none as measured.

What the 51 seconds can tell you without believing anyone

The video is the only evidence with any independence from the claims made about it, because it can be inspected. Three observations hold regardless of who made it or how.

Its length is a fraction of the demos it is being compared to. The strongest prior Claude video demo in the public record is a 156.6-second music video built with Claude Opus 5.5, published with its source code on GitHub, whose README states it took two generations rather than one. The demo that produced the "2.5 hours" claim is 51.4 seconds. A piece of work under a third the duration is not automatically less impressive — a tight 51 seconds can be harder than a baggy three minutes — but it is a different claim, and "2.5 hours" against 51 seconds is a number that should be read as this is what one session produced, not as a rate at which finished video comes out of a model.

"Existing assets" is doing a lot of descriptive work. The relaying post is explicit that the images and the music were not generated by the model. That means the model's contribution is the part around them: sequencing the stills, timing cuts to the audio, adding motion and transitions, drawing whatever overlays appear in frame, and writing the renderer that puts it all together. That is a real and useful capability, and it is the same capability the Opus 5.5 demos exercised. It is not the capability the phrase "image and video gen model obsolete" describes. A system that arranges existing images to existing music is not competing with an image generator or a video generator; it is competing with a video editor.

The claim that the model "chose the images" is the one to be most careful with. Choosing images from a set you were given, choosing music from a set you were given, and choosing a visual direction are three different levels of autonomy, and the public record here does not distinguish them. He could have been handed a folder and a track and asked to build something. He could have prompted for candidate assets and picked from the results of a generator. The post says "it was free to do its own research and make the creative decisions," which is a statement about how the session was framed, not about where the raw material came from. Until the source publishes the assets, the prompt or the process, "chose" is doing the work that a description should be doing.

None of that makes the demo dishonest. It makes it under-specified, which is the normal condition of a launch-eve test post and precisely why this piece is labelled the way it is.

Why this demo class already has a name in Anthropic's lineup

The claim in the video would be far stranger if Anthropic's shipped models had never done anything like it. They have, repeatedly, and the pattern is documented well enough to describe precisely.

Across the two days after Anthropic shipped Claude Opus 5.5 on September 22, 2026, several people published code-rendered films and animations built with it, and a few published their source. The pipeline is consistent every time: a page using p5.js, Canvas 2D or WebGL paints each frame; a Node script drives a headless Chrome instance and screenshots the canvas; ffmpeg joins the frames and lays the audio on top. At 24 frames per second, a 156.6-second video is roughly 3,760 screenshots taken one at a time. The model writes the renderer and directs it. It does not generate a single pixel on the inference call.

That distinction is the whole reason this class of demo keeps getting read as something it is not. "Claude coded a video" sounds like a claim about a generative capability. It is actually a claim about sustained code generation — the model holding a large, multi-file, multi-hour build together well enough that a second pass is a revision rather than a restart. The expensive part of the pipeline, frame 2,000 of 3,760, is a browser screenshot on somebody's laptop. No model call happens there.

There is an independent number that describes what that costs the model, and it is worth having in hand before weighing any claim about a future one. Artificial Analysis measured Claude Opus 5.5 on its Intelligence Index at 58 at maximum effort — the highest score it has recorded — 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 1.6× 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 its budget on itself. If a successor makes this class of demo cheaper or faster, this is the metric that would show it, and it has not been published for anything called Claude Next.

What would settle it, and what can be settled today

Five artifacts would move this from a claim into a fact, and they arrive in a fairly predictable order once an Anthropic model is real.

• A model ID. Something in the claude-{name}-{major} shape, or a variant that resolves on the Anthropic API and on the gateways that mirror it. Without one there is nothing to call, and no way for a second tester to reproduce anything.

• A rate card. Every shipped Claude model has published per-million-token pricing on the same day it shipped. A model that a tester can hammer for 2.5 hours of continuous generation already exists on some account's billing sheet.

• A row in the models overview. The vendor's own compare table is the two-minute check, and it is the one artifact that lists every shippable Claude model beside its API ID, its price, a 1M-token context window, 128K maximum output, and its thinking configuration. A model that reaches general availability gets a row there; a model with no row is not callable by anyone, whatever a post says about its quality.

• A Claude Code changelog line. Every model addition lands there with its identifier and price in the changelog message itself: Claude Opus 5.5 in version 2.1.280, Claude Haiku 5.5 in 2.1.293. It is the first place a new model becomes addressable by anyone outside Anthropic.

• The demo closed out. The source code, the asset list, the prompt, or simply the further tests @chetaslua said would follow. That is the artifact that would let someone else check whether 2.5 hours is a reproducible figure or a good session.

Two of those checks you can run yourself in an afternoon, and neither depends on a rumour being true. Point a client at a model string that your provider does not serve and watch what happens: Claude Code accepts a model string from /model during a session, the --model flag, the ANTHROPIC_MODEL environment variable and the settings file, and it does not validate the string up front when it comes from the flag, the environment or settings — it sends it and lets the request fail. That is exactly how a leaked name behaves if the name is real but the access is not yours: it looks like a typo rather than a permission denial. Then ask an endpoint what exists, because model lists are server-side facts. The Models API returns everything available to your key with its token limits and capability object. If it is not in that response and not in the docs table, no application code can reach it, whatever a post says about its quality.

The routing question this actually raises

Strip the enthusiasm and the interesting claim is not that an unreleased model can code a video. It is that a model which does not exist on any vendor surface was, reportedly, being tested through a client rather than through a documented API — and that the tester's evidence is a video rather than a benchmark. Both of those are normal for a pre-launch model, and both make the near-term decision for anyone reading this unusually concrete.

Anthropic's current shipped lineup is four models across a wide price range: Claude Haiku 5.5 from $0.10 per million input tokens and $0.50 per million output, Claude Sonnet 5.5 at $2 and $10, Claude Opus 5.5 at $4 and $20 with cache reads at $0.20 per million, and Claude Fable 5.1 at $10 and $50 for demanding reasoning and long-horizon agentic work. All four sit behind one OpenAI-compatible endpoint on OrcaRouter at provider list price with 0% markup, so a vendor price change is live here the same day rather than after a sync. If a Claude Next does ship and lands between two of those rows — the way Sonnet 5.5 landed between Sonnet 5 and Opus 5 — the question a reader will actually have is what it costs per task at the effort level that produces work like this, and that number does not exist yet for a model with no rate card.

That is also where the demo's real lesson sits, and it is a lesson about the shipped model, not the rumoured one. A 51-second video produced by writing a renderer is not evidence that video generation is solved. It is evidence that long-horizon code generation has become good enough to be a production strategy — and the way to test a claim like that against your own workload is to run it beside the models you already trust, on one key, with automatic failover so that an unproven route can degrade without taking your pipeline with it. Trying a demo-grade technique is cheap. Betting a production path on a rumour is not.

The open question

What the record supports, stated plainly: a 51.4-second video exists, posted by a tester who says it is "fully coded" and that the model behind it is called Claude Next; a second account relayed that claim 16 hours later, adding that the images and music were existing assets and the surrounding production was code, and attributing a 2.5-hour coding time to the tester; and nothing in Anthropic's documentation, changelog or rate card mentions the model at all. What the record does not support: that anything has been released, that any figure here was measured, or that a model named Claude Next makes an image or video generator obsolete. The same class of demo is already something Anthropic's shipped flagship does, with source published, at three times the duration.

So the honest position is the boring one. This is a leak-watch item, it is early, and it is the kind of early that is worth tracking rather than acting on. The thing to watch is not the next video. It is the first time the string shows up somewhere you can check: a changelog line, a row in the compare table, a model that answers when you call it, or a rate card with a number on it. When one of those appears, the demo stops being a claim about capability and becomes a line item you can route.