01Kimi vs Claude Opus: what this test compares
Kimi vs Claude Opus is a head-to-head test of Kimi K3.1 Pro and Claude Opus 5.5 on the same five creative coding builds. Claude Opus 5.5 won every round, while Kimi K3.1 Pro came out cheaper.
Julian Goldie ran the test to find out which model to use, what each costs, and what made the biggest difference. The five builds were a 3D open world, a neon racing game, a motion design title sequence, a cinematic movie trailer, and a living Japanese woodblock painting with a boatman rowing across a lake.
The rules were the same for both models:
- The same five briefs, word for word.
- One HTML file per build.
- No images, no 3D models and no stock assets.
- Everything generated in code, including terrain, water, cars and type.
Each model wrote a first draft, then got two rounds to improve it. Each round, the model saw a screenshot of its own build running, plus any errors and the frame rate. It was told to be its own harshest critic and fix every problem.
02Main differences between Kimi K3.1 Pro and Claude Opus 5.5
The two models were run in different places. Kimi K3.1 Pro ran through OpenRouter. Claude Opus 5.5 ran inside Claude desktop and Claude Code.
Neither model could see the other's work.
Here is how they compared across the whole test:
- Speed: Claude took 87 minutes for all five builds. Kimi took three and a half hours.
- Thinking time: Kimi sometimes thinks for 14 minutes before it writes a single line of code.
- Output: Kimi writes at around 33 tokens a second and is more verbose. That means more words and more waiting.
- Cost: Kimi's five builds cost $4.76 on OpenRouter. The Claude API would be about $8.
- Licence: Kimi K3.1 Pro is open source.
Julian also points out that the harness matters. Claude Opus 5.5 inside Claude desktop is designed to work that way. It feels more intuitive when you go back and forth with it.
03How to run a fair Kimi vs Claude Opus test yourself
You can copy the setup Julian used. The goal is to remove anything that could favor one model. Here are the steps from the episode.
- Write five briefs and give both models the exact same words.
- Set the same rules: one HTML file, no images, no 3D models, no stock assets.
- Run Kimi K3.1 Pro through OpenRouter.
- Run Claude Opus 5.5 inside Claude desktop in Claude Code.
- Ask each model for a first draft.
- Give each model two improvement rounds. Show it a screenshot of its build running, plus errors and the frame rate.
- Tell it to be its own harshest critic, fix every problem and make it better.
- Track every minute and every cent.
Julian says the exact prompts he used are inside his AI Profit Board community. That is also where he shares how he uses Claude Opus in Claude desktop every day for his business.
04Rounds 1 and 2: the 3D world and the racing game
Round one was an explorable 3D open world. The brief asked for hills, water, trees, fog, a day and night cycle, and something to discover.
Claude built a world at golden hour over a lake with real reflections. It had around 70,000 blades of grass moving in the wind, light rays through the trees and birds. You walk in first person and collect glowing shards to wake a shrine on the hill.
It then flows into night with the Milky Way and fireflies.
Kimi built Aurelia, a lake valley with a day and night time lapse and glowing echoes to find. It works, but it felt flatter and like an older generation of design. Claude won round one.
Round two was a futuristic anti-gravity racer. It needed a neon track, a city at night, five AI opponents, boost pads and three laps.
- Claude's Neon Rift had a giant sunset, glowing barriers, a live leaderboard, a minimap and a speedometer. It reads instantly.
- Kimi's Neon Velocity had a really nice HUD and a sharp race position design. The track was darker and harder to read at speed.
Both were playable from one prompt. Claude won round two.
05Rounds 3 to 5: title sequence, trailer and Edo Lake
Round three was a 30 second title sequence for a made-up AI product called Nova, with the tagline "think faster than thought." The target was a top studio look, like an Apple keynote opener. Julian cared most about this one, because a well-made short video helps any business that creates content.
- Claude: an orange orb breaks into particles, and the word lands huge on a cream background. Easing is smooth and spacing is clean.
- Kimi: bold blue and red scenes, a lime blob and scattering letters. Some frames look great.
Claude won, thanks to its attention to detail.
Round four was a 25 second cinematic trailer rendered live in the browser. Claude made After Light, with a gold title over a planet sunrise, a purple city, a figure in a red cape and a rocket lifting off. Kimi went graphic, with a giant eye, a rocket ignition and a nice timeline.
Julian did not love either, but Claude won a closer round.
Round five was the Edo Lake, set in Japan around 1820. Claude showed Mount Fuji, a red torii gate in the water and a standing boatman whose strokes leave ink ripples. Time labels were in Japanese.
Kimi's boatman sat down, moved slower and had duller colors. Its final round also broke the build with one typo.
06Speed, cost and final scores in the Kimi vs Claude Opus test
Claude Opus 5.5 won five rounds to zero. Julian says the design was smoother, the detail was better, it was more fun to watch, it was faster, and every build worked.
Here are the numbers:
- Claude time: 87 minutes for all five builds.
- Kimi time: three and a half hours, about 2.4 times slower.
- Kimi cost: $4.76 on OpenRouter, about 95 cents per build.
- Claude API cost: about $8, roughly double.
Julian ran Claude inside his Claude plan, so he had no extra bill. Cost is Kimi's win. It is also where Kimi's weakness shows: it is a bit slower, more verbose, and the design quality is not at Claude's level yet.
One more point: Kimi's last round broke the whole 3D screen. What Julian showed was Kimi's previous version.
07Tips, limits and mistakes to avoid
A few lessons come out of this test. They apply if you want to compare models or pick one for building.
- Give a fair brief. Use the same words and the same rules for both models.
- Use improvement rounds. Showing a model a screenshot, errors and frame rate helps it fix its own work.
- Watch for breakage. Kimi's last round broke because of one typo. Keep the previous working version.
- Plan for wait time. Kimi can think for 14 minutes before writing code.
- Count the real cost. The API price can differ from what you pay inside a plan.
- Expect to iterate. Julian says he would go back and forth a lot more on something like the trailer.
Remember that the harness matters too. Claude Opus 5.5 inside Claude desktop felt more intuitive for back-and-forth work.
08Who should pick Kimi or Claude Opus
Julian's pick is Claude Opus 5.5. He says it is his go-to, and it is the one he would choose if he had to pick just one.
Here is how to decide:
- Choose Claude Opus 5.5 if you want the best design, strong detail, faster builds and every build working. It also suits you if you already pay for a Claude plan.
- Choose Kimi K3.1 Pro if you are on a tight budget, you are not in a rush, and you want to use the API. It costs about half the price.
- Consider Kimi if open source matters to you.
This matters for anyone making content for a business. A polished title sequence or interactive page can look like a designer made it. Julian shares his prompts and daily Claude workflow inside the AI Profit Board community.