Which Claude Model Is Best for Motion Graphics? I Tested All Four

Claude Motion is only available on Teams and Enterprise plans right now. If you're on Pro or Max like me, you're waiting.

So I stopped waiting. I built my own tool, Clearmud Motion, where I describe an animation, Claude writes it, and it renders to a video file on my server. Then I ran one brief through all four of Anthropic's top models to see which one I'd trust with my videos and paid ads.

Short version: Sonnet 5.5 if you're on a budget, Opus 5.5 if you can pay a little more. The long version has a few surprises.

What Clearmud Motion is

Let me be clear about what this is and isn't. I built this for me. I use it to make motion graphics for our videos, transitions, and paid ads for clients.

It is not public facing, because I'm not paying API costs. It ties into my existing subscriptions and runs locally through my ClearMud HQ server. Right now it only toggles between GPT and Anthropic models.

Claude Motion is powered by the same models. So if you're locked out, my advice is simple: build your own motion tool instead of waiting. It's never been easier to build.

Test one: a UGC ad for Muse

The brief was a UGC-style ad for Muse, an app that lowers your bills. The headline: "My AI lowered my bill while I slept." The prompt template has a goal, a tone, a story, the copy to include, a final frame, and creative-freedom notes. I'll put the template in the YouTube description.

Settings for the run:

• Models: Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5
• Format: vertical 9:16
• Length: 30 seconds
• Blind test on, so I didn't know which output belonged to which model

The tool tracks output tokens, the equivalent API cost, and whether a model searched the web to complete the request. That turned out to matter.

The numbers before I watched anything

I checked the stats first [01:32].

| Stat | Result | |------|--------| | Slowest model | 570 seconds | | Fastest model | 282 seconds | | Cheapest run | $0.04 | | Most expensive run | $1.48 |

The part that surprised me: the second-fastest model was the most expensive. I expected speed and price to line up. They didn't.

Every model failed the logo

I had the prompt tell each model to go find Muse's brand assets on its own. All four came back with a wrong logo. Muse's mark is supposed to be an M, and nobody got there.

I think that's partly my prompt and not just the models' capabilities. But the lesson is the same either way. If you're doing this for a brand, or doing it to get paid, you attach the brand assets yourself. Don't make the model go hunting.

The reveal: Sonnet won

Before the reveal I was leaning between A and C. One of them had keyboard-typing text that I loved. I went with C as the winner and A as the runner-up.

C was Sonnet 5.5. A was Opus 5.5 [03:36].

That Sonnet pulled that off from one prompt, with zero refinement, is wild to me. If you're on a budget, Sonnet 5.5 is the way to go. For my own work I'm using Opus for nearly everything now, except research, which Fable 5.1 handles.

Haiku deserves credit too. It built a phone mockup with a text-message back-and-forth and nailed the outro. What I wanted was no black space around the phone. Zoom in, or use a light video overlay in the background so it doesn't overpower the shot. Simple, but not bad for the cheapest option.

The chat feature: letting Fable write the brief

I'm not an advertiser. I'm not a director. I don't always know how to explain what I want in an ad.

So I built a chat feature inside Motion. I can paste a prompt and generate, or I can pick a model and talk it through first. I like researching with Fable, so I used it for test two.

My rough ask was a 1:1 UGC ad for Grockbot, an inbox-sorting bot, that we could run on Facebook and Instagram, with the model pulling accurate brand assets like the logo, colors, and typography. Fable turned that into a master prompt with a shot list, tone, story, copy, and final frame [07:48].

The ad script it wrote:

> I gave a bot my inbox for a week. I named it Piper. One job. Sort my email. Flag what matters. It keeps going after I close my laptop. It only pings me when it needs a yes.

I set the run to 1:1 and 15 seconds and hit generate.

Test two: the cheapest model out-drew the most expensive

Model A was the most expensive in this round. Model B was the cheapest. And I was more impressed with model B's graphics than model A's [09:52]. Based on price I guessed B was Haiku.

Model A also failed to pull the branding, even though I asked for it. Same lesson as test one.

My winner was model D. It got closest to the logo. Overall this round was less impressive than the Muse ad, but still strong for a single prompt.

You never take the one prompt

This is where refinement comes in. Nobody ships the first output.

• Attach the brand logos and assets as files
• Add a slight pause between certain slides
• Fix spots where the title and the hook below it got read as one jarring sentence

Those are small fixes. The model did the heavy lifting.

Second reveal: Opus and Sonnet again

Opus 5.5 and Sonnet 5.5 landed on top again [11:23]. That's a trend across both tests.

Fable wasn't bad. But comparing Fable to Opus for motion graphics, I'll choose Opus 10 times out of 10.

In this round Opus cost double what Sonnet did and took about a minute and a half longer. Still, if you can bite the bullet, pay the little bit extra for Opus 5.5. If not, Sonnet 5.5 will work just fine.

Haiku could probably get there too with more creative direction and attached brand assets. The more we give these models, the better job they do.

What this test was really about

I wasn't only testing animation quality. I wanted to know if a model can go fetch the information I need so I can stay as hands-off as possible.

Based on both runs, they're not quite there yet. For Muse they made up their own logo. For Grockbot the most expensive model skipped the branding entirely.

Key takeaways

• Sonnet 5.5 on a budget, Opus 5.5 if you can pay a bit more
• Attach your brand assets yourself; don't rely on the model to find them
• Use a chat step to turn a rough idea into a full brief before generating
• Refine after the first pass; one prompt is a starting point, not a deliverable

I'm not an AI expert. I'm building in public and sharing what works. If you want a video on using Muse to pull from the Meta ads library through a Meta business account, drop a comment and let me know.

Watch the full test here: https://www.youtube.com/watch?v=oUUAyYNxsgY