Published: 2026-10-26
Analysis

Kimi K3.1 Pro vs Claude Opus 5.5 on five builds: Opus faster and better every time, Kimi two to four times cheaper

Chapters / key moments (click to jump — plays here on the page)

Kimi K3.1 Pro was released this week. Bart Slodyczka ran it against Claude Opus 5.5 on five builds, each model in its own native desktop app, and scored speed, cost and quality. The numbers below are his measurements from one run each, not a benchmark, and Moonshot's own documentation does not mention K3.1 Pro yet.

Source video

"I Tested Kimi K3.1 Pro vs Claude Opus 5.5 (RAW RESULTS)" by Bart Slodyczka — Watch on YouTube →

What is and isn't confirmed about K3.1 Pro

The video's account. K3.1 Pro is the same 2.8-trillion-parameter model as Kimi K3 with more reinforcement-learning training, and he says it scores three points below Opus 5.5 on the Artificial Analysis-style index he cites and about five times higher than K3 on task automation. These are his statements.
Moonshot's docs. The Kimi K3 quickstart (read 2026-10-26) documents only kimi-k3 with a 1M-token context window and does not mention K3.1 Pro, so we have no official model ID, price or release note to cite.
OpenRouter. It lists moonshotai/kimi-k3.1-pro, created 2026-10-23, with a 1,048,576-token context window and its lowest-priced endpoint at $2 per million input tokens and $10 output. That is a router listing, not Moonshot's price; see our model pricing page for how we treat router prices.

The five builds (his measurements)

Build	Kimi K3.1 Pro	Claude Opus 5.5	His verdict
3D boss-fight game (two prompts)	2h 08m, $9.67	1h 31m, $18.54	Opus: richer 3D and a harder fight; Kimi's boss died after about 10 arrows
3D apartment walk-through (two prompts)	3h 15m, 13.6M tokens, $6.60	56m, 31.9M tokens, $17.17	Opus: correct chair and toilet orientation and doors; Kimi had a light-switch win
Blender pickup truck	5h 30m, 10.4M tokens, $5.58 (one build)	1h 23m for two builds, 44M tokens, about $23	Opus: more detail and no colliding parts
Product-launch motion graphic (Three.js)	1h 14m, 2.4M tokens, $1.25	22m, 6.2M tokens, $4.34	Opus: more detail
Reverse-engineer ClickUp	2.5h, 13.8M tokens, $5.88	33m, 27M tokens, $14.03	Opus: closer to ClickUp; neither near 60 percent

Opus won on both speed and quality in every build, for 10 points. Kimi won on cost every time, at about one-half to one-quarter of Opus's cost per build.

Method and extras

Same prompt for both models, each run in its native harness: Kimi in the Kimi desktop app, Claude in the Claude Code desktop app. Costs are API prices as he calculated them; he says Kimi is roughly half the price of Opus per token.
For the ClickUp test he used the "reverse engineer anything" repository (linked from his earlier video). He says it is not something to run on every app.
A single Kimi follow-up: an Age of Empires II clone cost $1.19 for the first version (25 minutes, 1.6M tokens) and $2.68 more for a better-looking second version (14 minutes, 6.6M tokens); its 3D look was good, but units walked into buildings and could not path around trees. A soundtrack took 8 minutes and $0.45, and a two-minute music video cost $2.23 more.

Gotchas & Caveats

One run per model per build, with prompts he reused from earlier tests, so treat it as a field report, not a benchmark. He says as much about domains: these builds are not where Kimi is meant to be strongest, and he plans a follow-up on office automation.
He mentions a claim that Moonshot toned down some capabilities before a wider release in early November, and that Anthropic recently blocked a reverse-engineering repository from working with Claude. We could not verify either and have not relied on them.
Wall-clock times depend on the harness and server load. One Kimi truck build took 5.5 hours, and a music video was stuck for over an hour.

Key Takeaways

On these tasks the cheaper model was slower and needed more tokens, so compare cost per finished result, not per token.
Check Moonshot's own model page for K3.1 Pro's ID and price before building on a router listing.