Kimi K3.1 Pro: weights promised by November 4, price held at GPT-6.1 Sol level

On October 23, 2026, Moonshot posted Kimi K3.1 Pro to the Kimi API. There are no open weights — the company promises to publish them by November 4, 2026. Below is only what is in the announcement: access terms, prices, benchmarks and one manual check.

API today, weights with a promise by November 4

The model is available through the Kimi API (an HTTP interface through which a request goes to Moonshot servers and a response comes back). These are not open weights (open weights — when the model file can be downloaded and run on your own hardware). For now, only network calls.

Wording from the announcement: "Full model weights will be released by Nov 4, 2026". There is no data on exactly what will be included in the release and under what license.

A Highspeed version is offered separately. According to Moonshot, it works 2–3 times faster and costs twice as much. The exception is input token cache hits (input token cache: if the same piece of prompt has already been sent, the server may not recalculate it). For cache hits, the price is the same as for the regular version.

The mechanics are simple: speed is a resource that is sold separately. If you need the answer by a deadline, you take Highspeed and pay the double rate. If you can wait, the regular version is cheaper.

Price: $2.00 per million input, $10 per million output

The tariff in dollars is $2.00 per million input tokens and $10 per million output tokens. This is the same as GPT-6.1 Sol, and half the price of Claude Opus 5.5. For Highspeed, it is twice as expensive, except for input cache hits.

The asymmetry matters: output is five times more expensive than input. But input tokens are usually tens of times more numerous. Example: an instruction of 10 thousand tokens and an answer of 500. Input costs $0.02, output costs $0.005, so here the long instruction is four times more expensive than the short answer. Once the answer grows to 5 thousand tokens, it will cost $0.05 — 2.5 times more than the whole instruction. Therefore, in the budget count both numbers, not just the price per million.

Cache changes the picture if you run the same system prompt or request template many times. Then the repeated part of the input is not paid at the full rate.

Benchmarks: 55 at Artificial Analysis and a caveat about the vendor

Moonshot's own benchmarks (benchmark — a standard set of tests on which models are compared) put Kimi approximately at the level of Opus 5.5 and GPT-6 Astra: higher on some tests, close on others. These are the numbers of an interested party; they should be read as a claim, not as a measurement.

The independent assessment is Artificial Analysis, an aggregator that runs models through a set of tasks and reduces the result to one score. Kimi K3.1 Pro has 55 there. This is higher than Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon (all three have 53) and lower only than Claude Opus 5.5 (58).

The substance is not in the number itself but in the position: according to the primary source, this is the first case when a model with soon-to-be open weight status has risen so high against the frontier in this score. The caveat is obvious: a single-score aggregate does not show which tasks the model underperforms on. A difference of two points on such a scale is neither a verdict nor a guarantee.

How to check by hand: pelican, reasoning_effort and three levels

The author of the primary source ran the model through the Moonshot API with the command llm openai endpoint, which he added to LLM 0.32 in August. The command can send a prompt to any OpenAI-compatible endpoint, but does not log the result — therefore stdout and stderr (the reasoning trace goes to stderr, the trace of intermediate reasoning) were saved to separate files.

The flags that set behavior: --responses — use the Responses API; -o reasoning_effort low — the reasoning effort level; -m kimi-k3.1-pro — the model itself.

Reasoning effort is a handle that determines how many internal steps the model makes before the final answer. The model supports three levels: low, medium and max. More steps — longer generation and more output tokens, and output, as we remember, is the most expensive. Fewer steps — faster and cheaper, but on tasks that require a chain of calculations, quality may drop.

The test is the same prompt at all three levels: an SVG of a pelican on a bicycle. For comparison, the same three runs were made on Kimi K3. The images themselves are in the primary source; there is no textual assessment of how much the pelican on max differs from the pelican on low — only images.

Why this exercise. First, it checks that the model really responds through an OpenAI-compatible interface, not only through the native SDK. Second, it shows the spread between reasoning effort levels on a task where there is no single correct answer: the SVG is either drawn or not. If you have a similar task — generating markup, code, a diagram — start with low and raise the level only where the result consistently breaks.

What to do with this

1. Need access now — go through the Kimi API, the model has been there since October 23. There are no downloadable weights before November 4, and there is nothing on which to plan a local launch earlier than that date.
2. Count the budget by the input/output pair, not by one number. Output is five times more expensive than input, so long answers and high reasoning effort hit the bill harder than a long prompt.
3. If you run the same request template repeatedly, check the input token cache — on it the tariff does not double even in Highspeed.
4. Run your own "pelican": one real task on low, medium and max, saving stdout and stderr. This way you will see where the reasoning level really changes something and where it only adds tokens.
5. Treat the number 55 as a guide. It is higher than 53 for three models and lower than 58 for Opus 5.5, but it does not say on which of your tasks the difference will show. Until November 4, the only way to check is your own test through the API.