Can Sonnet make a better video than Opus? A blind test of 12 videos

People say Claude Opus 5.5 finally cracked motion graphics. Anthropic released Sonnet 5.5 six days after Opus. On several of Anthropic's own benchmarks, at max effort, it comes close to Opus, at half the standard price for input and output tokens. So I asked: can Sonnet make a better video than Opus?

I tested it blind. Below: both prompts, what each of the 12 videos cost, and what I found.

What I found

• Round 1, a detailed prompt: eight videos. The add-ons barely changed anything. To me they all looked about the same, and none wrote the word Claude in Claude's real wordmark. That was my prompt's fault: I asked for the word, not the wordmark.
• Round 2, a free-hands prompt: four videos, each a different idea. My blind picks were the two with no add-ons: one from Sonnet and one from Opus.
• The tokens: Opus used about twice as many as Sonnet for the videos I picked (16,023 against 7,440 output tokens).
• The add-ons: Motion and anime.js barely got used. Claude mostly wrote its own code.

What it cost

API list-price equivalents, one run per setup. Time is from start to finished MP4.

Anthropic's list prices are $2 and $10 per million input and output tokens for Sonnet 5.5, and $4 and $20 for Opus 5.5 (retrieved 4 Oct 2026). Opus costs twice as much per token. In these runs it cost about three times as much per video, because it wrote more.

Round 1: a detailed prompt

The same prompt for all eight videos, with a bit more installed each time. Press play on any clip.

Make a 10-second 2D motion graphic, 1440×1080 (4:3), and render it to an MP4 named claude.mp4 in this folder.
Style: vector-flat, minimalist. Palette: Claude's own colours only: Claude orange #D97757 and #DC6038, ivory #FAF9F5, on matte black.
0–3s: A flat Claude-orange square sits on a dark dot grid. It shatters into hundreds of small glowing orange dots that burst outward, then snap back together into the word "CLAUDE".
3–7s: One letter at a time, each letter morphs into a different shape (circle, triangle, wave) and back, on a staggered rhythm, while the grid scrolls slowly underneath.
7–10s: The letters stretch and flow upward into one glowing orb at the centre. The orb pulses three times, then snaps into the Claude logo, exactly as drawn in claude-logo.svg in this folder. End on a clean, still frame.

Nothing added

Sonnet wrote its own video renderer, in Swift. Opus built a web page and captured it frame by frame with the Chrome that came with a tool already on my Mac.

Plus Motion

Sonnet built a web page and never touched Motion. Opus loaded Motion, but only for easing curves, which set the timing of each move.

Plus Motion and anime.js

Neither used either library. Sonnet went back to its own Swift renderer, and Opus wrote a web page with its own capture script.

Plus Motion, anime.js and HyperFrames

Both loaded HyperFrames' instructions and rendered with it. Opus also used anime.js, as a timer. Sonnet used neither library.

Round 2: free hands

I changed the prompt: the real Claude wordmark as a file, and complete creative freedom. I ran two setups.

Make a 10-second motion graphic, 1440×1080 (4:3), and render it to an MP4 named claude.mp4 in this folder.
Its job: make someone who watches it want to use Claude, through the craft of the motion graphic itself. You have complete freedom over the concept, style, pacing and motion.
Brand rules: use Claude's real marks exactly as supplied, without redrawing or changing them: claude-wordmark.png is the wordmark, claude-logo.svg is the app icon. Use only Claude's colours: #D97757, #DC6038, ivory #FAF9F5, and near-black.
Don't put any claims, numbers or features about Claude on screen. The motion is the proof.

Nothing added (my blind picks)

Both wrote their own video renderers. These are the two I picked when I rated all four blind.

Plus all three add-ons

Both followed HyperFrames' instructions. Neither used Motion or anime.js.

The add-ons

I added them in this order, each run keeping the earlier ones, and told Claude only what was installed:

1. Motion (motion.dev, MIT): an animation library for the web. github.com/motiondivision/motion
2. anime.js (MIT): another animation library. github.com/juliangarnier/anime
3. HyperFrames (Apache-2.0): a free tool that turns a web page into a video, with instructions Claude reads before it builds. github.com/heygen-com/hyperframes

Run your own test

1. Pick one real job and write the prompt once.
2. Run it on Sonnet and on Opus, same effort setting, in an empty folder.
3. Judge the results blind: shuffle them and hide which is which.
4. Compare what each used. Your usage page shows tokens.

How I tested

Claude Code, Sonnet 5.5 and Opus 5.5, medium effort, 2 Oct 2026. Each run started in an empty folder with only the prompt, the logo files and a one-line note about what was installed. Round 1 was 4 setups × 2 models; round 2 was 2 setups × 2 models. I rated the clips blind: shuffled, with the model and setup hidden. Costs are API list-price equivalents from each run's own log. My picks and the costs show what happened in this setup, not a rule for every task.