Musk shares Opus 5.5 code-generated music videos as AI video challenge goes viral

Opus 5.5 generating music videos directly from code went viral in the community and sparked a same-prompt chain challenge. The core of the format is fixing the variables: the same song, the same reference video, handed to Opus 5.5 to reshoot entirely.

The one who started the chain is blogger donald. He threw the original link and audio of "Claude Pop," made by NotinReality, to Opus 5.5 together, requiring the song be kept and all visuals torn down and redone. Opus 5.5 then wrote code, made storyboards, matched visuals, and added transitions on its own, working for 12 hours to produce a 2-minute-21-second new MV. donald also publicly shared the complete prompt of about 9,500 characters.

The chain was quickly picked up by others. Pleometric turned the same challenge into an anime-style tech-apocalypse theme song, featuring a red-haired protagonist, a pixel stage, and retro game UI; mexicat switched to a more vivid and playful visual direction and did not pursue a "one-shot" result, instead taking the finished video and continuing to have Claude fine-tune it; anabology added Midjourney and a moodboard on top of donald's approach, letting Opus 5.5 work another 12 hours to make a super-long MV of over five minutes.

It should be noted that the finished quality, time spent, and duration of the above cases all come from the bloggers' own accounts, with no independent verification. What they can prove is that "someone did this and publicly shared the result," not that "anyone can reproduce the same result."

What truly pushed this chain challenge into the mainstream was Musk's reposts. He reposted the related MVs multiple times and gave evaluations such as "Whoa!" "Epic!" and "This time I truly felt AGI."

These evaluations are relayed by the source party, and the specific repost times and full context are not given in the available material. Therefore, what can be confirmed is the reported fact that "Musk reposted and gave such evaluations," and one cannot infer from this that he made any systematic judgment about Opus 5.5 or Anthropic.

From the perspective of the spread structure, a same-prompt challenge that originally happened among bloggers gained exposure far beyond the tech circle because of one high-attention repost—this is an inference of this article, based on the article narrating Musk's repost as the opening ignition point.

When the chain challenge was at its hottest, the opposing voice came from within. Claude Code team member Thariq complained that the post says Claude did it in one shot, but the prompt is 10,000 characters and also relies on skills, examples, and API keys.

Here a discrepancy in wording appears: donald's public prompt is about 9,500 characters, while Thariq says 10,000 characters. The two are close in magnitude but the numbers are not consistent, and the available material does not explain the source of the difference, so they can only be presented side by side, and neither can be taken as definitive.

The key point Thariq raised is not the character count itself, but what the phrase "one-shot" omits: skills, examples, and API keys are all preconditions outside the prompt. In other words, the omitted parts are precisely the parts that make the workflow run.

Blogger Movez broke down the entire workflow and concluded that the prompt accounts for only 10%, while the other 90% is the production framework. This ratio is Movez's personal judgment, not a verifiable measurement result, but the direction it points to is consistent with Thariq's complaint: what determines success or failure is what lies outside the prompt.

The motion techniques Movez summarized include: first feed references to generate a unified style guide, rather than just saying "make it more high-end"; make the animation into a seek(t) time function, so that by inputting the second, the frame for that second can be calculated, thereby supporting the browser to render different segments in parallel, and if one shot fails, the entire video does not need to be rerun; for an MV with a song, first analyze BPM, downbeats, and volume peaks to make a "beat map," then tie cuts, subtitle pops, color changes, and sound effects to these time points; finally, do not render the final version on the first try—first look at the storyboard and a low-resolution preview, arrange keyframes into a large image to check text, composition, and motion, pick out the three worst problems, fix them, and then render.

There are also specific numbers on the cost side: one of Movez's watercolor shorts called the model 163 times and ran for nearly 7 hours from start to finish. This number explains the distance between "one-click video" and the actual workflow better than any adjective.

In addition, running this workflow also has environment constraints: you need to select Opus 5.5 in Claude Code, and have Node.js and FFmpeg installed locally; when images and music are needed, you also need to connect other tools such as Midjourney and Suno.

Putting the above chain of evidence together, the article's conclusion is: "code-to-video" is real, but "one-click video" is not realistic. This is an inference based on case accounts and practitioners' rebuttals, not a conclusion verified independently.

What supports "code-to-video" being real is the finished videos and workflow descriptions publicly shared by multiple bloggers; what supports "one-click video is unrealistic" is the reliance on skills, examples, and API keys pointed out by Thariq, as well as the 10%/90% split and the measured numbers of 163 calls and nearly 7 hours given by Movez.

For readers, a more practical way to understand it may be: what Opus 5.5 lowers is the barrier from "idea to runnable code," not the barrier from "idea to finished video." The latter still requires reference material, beat analysis, low-resolution previews, and multiple rounds of rework—there is currently no evidence that these steps can be replaced by a one-sentence prompt.

The article also includes a leak. The article says that when Opus 5.5 was released, Anthropic planted an easter egg: Sonnet 5.5 and Haiku 5.5 will be launched in the coming weeks; Sonnet 5.5 is said to be released next week, speculated to be local time Tuesday.

The leak account JAZII released a set of same-prompt comparisons of four models: GPT-6 Astra, GPT-6 Sol, Claude Opus 5.5, and the allegedly secretly tested Claude Sonnet 5.5, simultaneously generating a pixel animation composed of a moonlit night, cherry blossoms, and a Japanese-style town. According to the leaker's description, the Sonnet 5.5 panel has denser architectural layers, richer foreground and background, and is more willing to pile on lighting, cherry blossoms, and small elements; the two OpenAI models have relatively flatter compositions, while Opus 5.5 is more restrained and clean. The leaker commented that this time Sonnet pressed down both OpenAI models.

This content must be treated as a leak: whether Sonnet 5.5 is really being secretly tested, and whether it beat the two OpenAI models, cannot be verified; the release timing is only "said to be next week," without official confirmation. The comparison result is only the leaker's evaluation and lacks first-party verification.