AI Motion Graphics Work Better as Code. Write the Animation, Then Hand It to a Video Model.
Ask a video model for a bar chart that grows to 42 percent and you will get a beautiful bar chart that grows to something. Maybe 40. Maybe 47. The label might read "42%" while the bar stops at a third of the way up. The camera might drift. The type will be close to your brand font and not quite it.
That is not a bug that the next version fixes. A video model paints frames that look right. It does not know that a number is a number.
Now ask a language model the same question, and ask it to answer in code instead of pixels. You get an animation where the bar lands on 42 because the code says 42, the label sits in your actual font, and the title hits on frame 30 because frame 30 is written down. Change the number and render again. Nothing drifts.
On October 8, two things made that second path much more practical for anyone who builds explainers, product spots, social cutdowns or title cards. The interesting part is what happens when you join them.
Two releases, one idea
The first is a benchmark. Cliphouse Research gave ten language models the same three motion-graphics briefs (a product launch, a piece of moving typography, and a data story) and asked each to write the animation in Remotion, a framework that turns code into rendered video (free for individuals and small teams). The brief called for a twelve-second, silent 1080p clip at 30 frames per second.
In the product-launch run the page publishes, eight of ten models produced a working video on the first try. The cost spread is what stops you. Claude Opus 5.5 passed for $0.374. Claude Haiku 5.5 passed for $0.011. GPT-6 Luna passed for less than a cent.
Those are the costs of writing the code, not of rendering it, and the benchmark only checks that the video compiled and rendered. It says nothing about which one looks good. The page says two runs are "exploratory evidence, not a statistically definitive ranking." Still, the floor has moved. A working animated draft from a cheap model now costs about what a single generated image used to.
The second release is a handoff. Luma announced that animations made in Claude Motion, a beta feature on Claude Team and Enterprise plans that writes animations as code, now open directly in Luma as video. From there, Luma's agent can restyle the look while keeping the motion, reframe it to 9:16, 1:1, 4:3 or 21:9, or make more versions.
Put those together and you get a workflow that plays to each tool's strength. The language model handles everything that must be exact. The video model handles everything that should feel alive.
Why the split works
Think about how a motion designer already works. The layout, the timing, the data and the type live in a project file you can reopen. The texture, the grain, the camera feel and the look get layered on top. Nobody would accept a tool that baked all of that into a single flat render and made you start over to change a date.
Generated video does exactly that. Every change is a reroll, and every reroll can move things you did not ask it to move.
Code reverses the problem. The animation becomes a text file. You can read it, change "42" to "44," swap a hex color, push a title three frames later, and render the same piece again. The file is your master, the same way a layered project file is.
The video model then becomes a finishing pass, the way a colorist or a texture pass is. If it misbehaves, you have lost a finish, not the piece.
Put this into practice
You do not need Claude Motion to work this way. You need any AI assistant that writes code, and a free way to render it.
Step 1: Start from a template rather than a blank page. Cliphouse offers free motion-graphics templates built for AI assistants. You can edit text and colors in the browser with no account, no watermark and no attribution required. A template gives the assistant a working structure to change, which is much more reliable than asking it to invent one.
Step 2: Write the brief like a motion designer, not a prompter. Give the assistant the facts it cannot guess: the exact numbers, the exact copy, your hex colors, your font, the clip length, the frame rate, and the beats. "Title in at 0.5 seconds, bar grows from 0 to 42 percent between 2 and 4 seconds, logo holds from 10 to 12 seconds." Timing written in seconds beats adjectives every time.
Step 3: Start with the cheap model. On the benchmark's published run, the inexpensive models wrote working videos too. Try one first. Move up to a more expensive model only when the cheap one fails to compile twice or produces something structurally wrong.
Step 4: Render and check the facts first. Before you look at style, check the things the code was supposed to guarantee: numbers, spelling, brand colors, timing against your script. If any are wrong, fix them in the code, not in the render.
Step 5: Save the code as the master. Name it like a project file. Every later version, a new language, a new date, a new figure, starts here.
Step 6: Hand the render to a video model for the finish. If you are on a Claude Team or Enterprise plan, the Luma connector does this directly. Otherwise, upload the rendered file to any video model that accepts a video in and ask for a style pass that keeps motion and layout. Ask for your social sizes here too.
Step 7: Check the finish against the master. Play the finished version next to the clean render. Did the generated pass soften your type, nudge a number, or change a color you were contractually bound to? If it did, either dial the style pass down or ship the clean render.
Where it breaks
The handoff is one-way. Once your animation is in a video model, it is a video. Change a figure and you go back to the code, re-render and re-finish. Budget for that loop.
A video model can still undo your precision. The restyle runs over your frames. If it is strong, it can blur small type or shift a number's shape, which is exactly what you went to code to avoid. Use the lightest finish that gets the look.
The benchmark measures "it ran," not "it's good." Passing means the code compiled and rendered. None of the published results judge design quality, and only one brief's first run is shown in full on the page. Do not choose a model on that table alone.
Remotion has a license line. The framework is free for individuals, non-profits, and companies with up to three employees. A for-profit with more than three needs a paid company license. Check before a studio adopts it.
The direct Claude-to-Luma path is limited. Claude Motion is in beta for Team and Enterprise only, and Luma's announcement does not mention cost. Luma's own plans start at $30 a month for its Plus tier, which includes commercial use. Restyling and reframing draw on those credits.
Templates are a starting point, not a license. Cliphouse says its templates are free to edit and remix with no attribution, but it does not publish formal license terms or list export formats. If a client needs paperwork, ask first.
Code has a ceiling. Charts, type, shapes and layout are code's home ground. Organic motion, a hand-drawn feel, anything that looks filmed: that is still where a video model, or a human animator, earns its keep.
The shift worth noticing
For two years, AI video has been sold as "type a sentence, get a finished shot." For motion graphics that promise has always been slightly wrong, because motion graphics are mostly about precision. The number has to be right. The logo has to be the logo.
What changed this week is that precision got cheap. A model that costs about a cent can write a working twelve-second animation, and a video model can now take that animation as its starting point instead of a blank prompt.
So try one piece this way. Take an explainer you would normally build by hand, write the beats in seconds, and let a cheap model write the first version as code. Fix the facts in the code. Then decide whether it even needs a generated finish. You may find the clean render was what the client wanted all along.
Medium metadata
Title: AI Motion Graphics Work Better as Code. Write the Animation, Then Hand It to a Video Model.
Subtitle: A new benchmark shows a cheap AI model can write a working motion-graphics clip for about a cent, and Luma now accepts Claude-written animations as video. Here is the code-first, finish-last workflow that keeps your numbers right.
Tags: Motion Design, Generative AI, Video Production, Animation, AI Tools
Estimated read time: 7 minutes