Can Claude Create Videos? What It Can and Cannot Do
On this page
Documentation checked September 16, 2026: Anthropic lists documents and still images for chat uploads; its published upload list does not include video formats. This describes the documented chat upload interface, not tools or integrations.
The Claude vision API accepts GIF images but processes only the first frame of an animation.
What Claude Makes for a Video
Split production into stages and Claude covers most of them well. It just never touches the pixels.
| Task | Claude alone | With tools | Notes |
|---|---|---|---|
| Write a script | Yes | Yes | Its strongest contribution |
| Storyboard and shot list | Yes | Yes | Text-based, scene by scene |
| Per-scene video prompts | Yes | Yes | For a generator to execute |
| Captions and metadata | Yes | Yes | Needs a transcript to work from |
| Code-based animation | Partly | Yes | SVG and CSS render in Artifacts |
| Render an MP4 | No | Yes | Remotion or FFmpeg does the render |
| Generate footage from a prompt | No | Yes | Needs a text-to-video model |
| Edit existing footage | No | Via code | Cannot see what it is cutting |
The pattern is the same one that shows up with every language model and video: strong on anything expressible as text, absent on anything that requires seeing.
Can Claude Create Animations?
Yes, and this is the part people miss. It just is not video.
Ask Claude for an animated SVG, a CSS keyframe animation, or a React component with motion, and it writes working code that renders in an Artifact. You get real movement on screen. It looks like animation because it is animation.
The distinction: Claude wrote instructions your browser executes. It did not produce frames. Nothing about that output is a video file, and you cannot upload it to YouTube without a separate render step.
That is genuinely useful for explainer graphics, chart animations, UI motion prototypes, and loading states. It is not useful when you need footage of a person, a place, or a product. The same split shows up in whether ChatGPT can create videos, where the answer changed after Sora shut down.
Four Ways to Get an Actual Video
Ordered by how much code you want to write.
Claude writes, a generator renders
No code. Ask for a scene table with one generation prompt per shot, take those prompts to a video generator, assemble the clips. The prompts are where quality comes from, which is exactly the part Claude is good at. The roundup of free AI video generators covers which ones to feed them into.
Claude Code plus Remotion
Remotion builds video from React components, so a video becomes a codebase. Claude Code is unusually good here because the task is writing React, not making pictures. Best fit is templated output: the same branded intro over 200 personalised endings, data-driven charts, release notes as motion graphics.
Claude Code plus FFmpeg
For manipulating footage you already have: concatenating, trimming, transcoding, burning in subtitles, batch resizing. FFmpeg syntax is notoriously fiddly and Claude writes it correctly more often than most people do from memory. The catch is that it cannot watch the result, so you review every output yourself.
Connect a video service by API or MCP
Claude orchestrates: calls a text-to-video or avatar service, passes prompts, polls for the render, handles the files. Worth the setup only when you are producing at volume. For one video it is slower than opening the tool yourself. Connectors work in the other direction too: ScreenApp's MCP connector lets Claude read videos you already have.
A prompt worth reusing for the first route:
"Break this 30-second product video into six five-second scenes. For each: the visual subject, the action, the environment, camera movement, lighting, the voiceover line, on-screen text, and a self-contained text-to-video prompt I can paste into a generator. Keep the product description identical across all six prompts so it does not drift between shots."
That last instruction about keeping the description identical is the one that matters. Reworded product descriptions between scenes are the main reason AI-generated sequences look inconsistent.
Can Claude Code Make Videos?
It can produce one, but it is not generating it. The distinction holds all the way down.
Worth checking your plan before committing to this route, since code execution and file creation have rolled out tier by tier. The Claude pricing breakdown covers what each tier includes.
Claude Code writes the render script, installs the dependencies, runs the commands, and hands you a file. Remotion or FFmpeg does the actual rendering. Take those away and Claude Code produces nothing playable, because the ability was never in the model.
Where this genuinely pays off is volume and repetition. One video by hand is faster in a normal editor. Two hundred videos with swapped data, or a nightly render triggered by a build, is a job only the code route can do.
Which Tool for Which Job
Three different products, three different jobs, and picking wrong is why people conclude AI video is disappointing.
| You want to | Reach for | Why |
|---|---|---|
| Plan and script it | Claude | Structure and language are its strengths |
| Automate 200 of them | Claude Code plus Remotion | Video as code scales, hand editing does not |
| Invent footage that never existed | A text-to-video generator | Only a video model makes new frames |
| Cut footage you already have | A real editor | You need to see what you are cutting |
| Pull clips from a long recording | A clip generator | Finds the moments without a code step |
If you already have recordings, the code route is almost always the wrong answer. An AI video editor cuts footage directly, a clip generator pulls short segments out of a long recording, and a video analyzer tells you what is in it. All three do things Claude structurally cannot, because they can see the frames.
Where the two combine well: get a transcript and an AI video summary of the recording first, then hand that text to Claude to restructure into scripts, posts, and chapter markers. Text is the handoff format.
What I Would Actually Do
Need one video and you have no footage: use Claude for the script and the per-scene prompts, then paste those into a generator. Twenty minutes of planning beats an hour of regenerating vague clips.
Need the same video 200 times with different data: Claude Code and Remotion, and it is worth the setup. This is the only case where the code route clearly wins.
Already have the recording: skip Claude for the video work entirely. Edit it in something that can see it, then bring the transcript to Claude for the writing.
