Moonshot releases Kimi K3.1 Pro: 55 points and $2/$10 per million tokens

China's Moonshot has released the Kimi K3.1 Pro model in its API. Pricing is $2 per million input tokens and $10 per million output tokens. GPT-6.1 Sol costs the same, while Claude Opus 5.5 costs twice as much. On the independent Artificial Analysis index, Kimi scored 55 points and trailed only Opus 5.5, which scored 58. The weights are closed for now, but Moonshot promises to publish them in full by November 4, 2026.

What the independent measurement showed

Moonshot's own benchmarks put K3.1 Pro roughly on par with Opus 5.5 and GPT-6 Astra. In some tests it pulls ahead of them, in others it comes close. Vendors tend to pick the tables that flatter them, so an external index matters more. Artificial Analysis ranked the models like this:

• Claude Opus 5.5: 58 points
• Kimi K3.1 Pro: 55
• Claude Fable 5.1: 53
• GPT-6 Astra: 53
• Gemini 4 Argon: 53

The gap to the leader is three points. Kimi sits two points ahead of the three closed models. Until now, open-weights models, even if the weights are still to come, have trailed the closed frontier by a clear margin. This time, for the first time, that did not hold.

Token economics and the price of speed

For agentic workloads, token price matters more than a difference of a couple of index points. An agent running a loop of "read the repository, think, fix" burns a lot of input tokens and noticeably fewer output tokens. Take Moonshot's rates: a session with 500 thousand input tokens and 50 thousand output tokens costs $1.00 for input and $0.50 for output, or $1.50 in total. With Opus 5.5, priced twice as high, the same volume comes to about $3. That is our own calculation, not vendor data. Across thousands of sessions a day, the difference becomes a noticeable line in the budget.

There is also a Highspeed variant. It is 2–3 times faster and twice the price. Input cache hits are the exception: their price does not change. For long agentic chains with a large repeating context (a system prompt, a codebase), that is a favorable condition. You can switch the fast mode on for interactive tasks and leave background runs on the standard tier.

Pelican on a bicycle and a test through LLM 0.32

Simon Willison tested the model with his usual benchmark: he asked it to generate an SVG of a pelican riding a bicycle. He ran it against the Moonshot API with the `llm openai endpoint` command from LLM 0.32, which he added in August. The command works with any OpenAI-compatible endpoint, in this case https://api.moonshot.ai/v1 with the model kimi-k3.1-pro. The command does not log results itself, so he had to redirect the output and the reasoning trace to stdout and stderr files. He used the `--responses` flag and the `reasoning_effort` option.

The model supports three reasoning levels: low, medium and max. Willison ran all three and, for comparison, reran the previous Kimi K3. He called the result at max fairly solid: the feet are on the pedals, the bicycle frame is well drawn, and the beak came out well. Among the flaws, both legs ended up on the same side of the frame.

What to check before November 4

Only the API is available for now, and the easiest way to test it is through an OpenAI-compatible client. You only need to change the base URL and the model name. To evaluate the model, run all three reasoning levels on your own tasks and compare not only quality but also output token consumption: at max, the reasoning trace noticeably eats into the budget. Separately, measure your cache hit rate, because the Highspeed price depends on it.

Teams considering self-hosting need to wait for the actual weights promised by November 4, 2026. It is worth assessing GPU requirements and license terms in advance. If the promise is kept, closed providers will gain a competitor with similar quality that can also run on your own infrastructure.