Video agent · Claude Opus 5.5
The Creative Claw video agent runs Claude Opus 5.5, the best model for motion design we have used, with Claude Code in the cloud. Ask in ChatGPT, Codex or Claude for a launch film, explainer, promo or data story. It writes the video, makes the media it needs, checks every scene and renders the MP4.
Each request starts a cloud machine running Claude Code on Claude Opus 5.5. The machine has a browser, a renderer and the HyperFrames tools, so the agent can do what a motion designer does: plan the scenes, write the animation, look at its own frames and fix what is off.
Every effort level uses the same model. Light, Medium and High set how long it thinks before and while it works.
One request, one finished video. Each request is called a turn.
Say what it shows, the exact words on screen, its length and its size. Attach footage, images, music or a logo if you have them, and say what each is for.
It writes the video as a composition of scenes, type and motion, and generates the images, voiceover, music and sound effects the brief calls for, inside the media budget you set.
The agent looks at frames from each scene and each transition, fixes what it finds, and the service renders the MP4 with sound at the size you asked for.
Say what to change in the chat, or use Request changes on the video card. The agent revises the same project and renders it again. Earlier versions stay in your library.
Both videos below came straight from the agent. Nothing was edited afterwards.
The request
“An 8-second vertical event teaser with two scenes: the headline SUMMER NIGHTS with the date, then FREE ENTRY BEFORE 23:00 and the web address. Dark background, one warm accent colour, bold condensed type.”
Medium effort · 89 credits · about 3 minutes
The change
“Change the date line to 31 JULY · SATURDAY and make the scene 2 headline about 15% larger. Change nothing else.”
36 credits · about 1 minute
Mention Creative Claw and describe the video. Give the exact words for the screen, the length and the size. These four requests cover the common cases.
Make a 15-second vertical launch teaser, 1080×1920, for a note-taking app called Quill. Three scenes: the problem in one line, the app name with the line "Notes that organise themselves", then "Free on iOS · quill.example". Dark background, one warm accent colour, bold condensed type, music that builds.
Use my attached video and logo. Cut a 20-second vertical teaser from the best moments, keep the original music from the video, add the date "24 October" in the middle and end on the logo with "Tickets at example.com". Do not generate any media.
Start from the Swiss Grid HyperFrames example and turn it into a 20-second data story about our third quarter: revenue up 42%, 3,100 new customers, and churn down to 1.8%. Keep its layout and motion.
Change the date line to "31 July · Saturday" and make the scene 2 headline about 15% larger. Change nothing else.
Most videos cost between about 25 and 450 credits ($0.25 to $4.50). You pay for the work the agent actually does, plus any media it generates.
Where a video lands depends on its length and on how complex the task is: more scenes, footage to cut, media to generate and extra polish all mean more work. Each effort level also has a hard ceiling, and you are never charged above it.
| Effort | Use it for | Most you pay for the agent's work |
|---|---|---|
| Light | Simple changes and basic videos. Fastest and cheapest. | 400 credits ($4) |
| Medium | A good fit for most videos. This is the default. | 600 credits ($6) |
| High | Complex or polished videos. Slower, and it costs more. | 1,000 credits ($10) |
What real turns cost
| Video | Effort | Credits | Time |
|---|---|---|---|
| A 6-second title card | Light | 25 | under 1 minute |
| An 8-second, two-scene event teaser | Medium | 89 | about 3 minutes |
| A small change to that teaser | Medium | 36 | about 1 minute |
Measured on our own account on 8 October 2026, with no generated media. Longer videos, footage to cut and generated media cost more: our heaviest turns so far, 25-second videos cut from supplied footage, cost between 300 and 450 credits. Credits are $10 per 1,000. Full pricing.
You do not have to start from a blank page. Name one of our 28 HyperFrames example projects and say what to change, and the agent builds on its layout and motion. Every video you make also comes with its own project file, which a later video can start from.
HyperFrames @heygen-com on GitHub
HyperFrames
View project
HyperFrames @heygen-com on GitHub
HyperFrames
View project
HyperFrames @heygen-com on GitHub
HyperFrames
View project
HyperFrames @heygen-com on GitHub
HyperFrames
View project
HyperFrames @heygen-com on GitHub
HyperFrames
View project
HyperFrames @heygen-com on GitHub
HyperFrames
View projectNames, dates, prices and web addresses appear exactly as you write them. Say which text belongs to which scene.
“Use the music from this video”, “this poster is the final frame”, “this clip is a style reference only”. The agent follows the roles you give.
The agent designs and cuts; it does not film people. For a moving portrait, generate a short clip with a video model first and attach it.
A focused change is quick and cheap. List the changes you want in one request and say what must stay as it is.
It is an agent that makes designed videos from a description. You say what the video should show, how long it is and what size. The agent writes the composition, generates the images, voiceover, music and sound effects the brief calls for, checks every scene, and renders a finished MP4 with sound.
A video model films a scene: people, places and camera moves. The video agent designs one: titles, layouts, charts, product shots, captions and motion, with exact text and exact timing. Use the agent for explainers, promos, launch teasers, kinetic type and data stories. Use a video model for photorealistic footage and character acting.
Most videos cost between about 25 and 450 credits ($0.25 to $4.50). Where a video lands depends on its length and on how complex the task is: more scenes, footage to cut, media to generate and extra polish all mean more work for the agent. You are charged for the work actually done.
The agent's own work is never charged above 400 credits at Light, 600 at Medium and 1,000 at High, whatever it took. Media the agent generates is added at its normal credit price and cannot exceed the media budget, which is 200 credits unless you set another. Most videos cost far less than these ceilings.
Every effort level runs on Claude Opus 5.5, driving Claude Code on a cloud machine that has a browser, a renderer and the HyperFrames tools. Opus 5.5 is the best model for motion design we have used. The effort level sets how long it thinks, not which model it is.
The final cost is only known when the agent finishes, so 200 credits are set aside at the start. If the video costs less, the difference comes back. If it costs more, the rest is charged then, up to the ceiling for the effort level. The account needs 200 credits available to start.
A turn that fails is not charged and the full hold comes back. If you stop a turn yourself, the work the agent had already done is charged, the rest of the hold comes back, and no video is made. The card asks you to confirm before it stops.
A small change takes about a minute. A short text-led video takes one to four minutes. Our heaviest turns so far, 25-second videos cut from supplied footage, took 16 to 23 minutes and cost between 300 and 450 credits. You can ask for an email when the video is ready.
Yes. Attach up to 20 files and say what each one is for. The agent can cut your footage, keep your original music, place your logo and match your poster. Set the media budget to 0 if you want it to use only what you supplied.
Yes. Every finished video comes with its project file, which you can start a new video from. You can also start from one of the HyperFrames example projects on our examples page: name the example and say what to change.
The agent writes videos in HyperFrames, an open-source format for video written as HTML.
Install Creative Claw in ChatGPT, Codex or Claude, describe the video, and the agent does the rest.