Kimi K3.1 Pro is out: closer to Opus on coding, a third cheaper than K3
Moonshot AI released Kimi K3.1 Pro on October 23, a post-trained version of its open K3 model, in the Kimi apps, Kimi Code and the API. It costs a third less than K3 per token, Moonshot's own tests put it within a point of Claude Opus 5.5 on terminal coding, and the weights are due by November 4.
What happened
Moonshot AI published Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed on October 23, 2026, and made both available the same day. K3.1 Pro is not a new base model: Moonshot describes it as Kimi K3 with much more post-training, reinforcement learning across hundreds of task domains at once. Highspeed is the same model served two to three times faster.
It is in the Kimi web app, the iOS, Android and HarmonyOS apps, Kimi Work 3.1.0 and later on Windows and Apple-silicon Macs, and Kimi Code, where you switch to it with /model. In the API the names are kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Business and Enterprise workspaces get it switched off until an admin turns it on.
The numbers Moonshot gives
From the launch post, all measured by Moonshot at maximum reasoning effort:
| Test | Kimi K3 | K3.1 Pro | For comparison |
|---|---|---|---|
| Terminal-Bench 4.0, Claude Code harness | — | 65.7% | Claude Opus 5.5 66.4%, GPT-6 Astra 58.2% |
| Kimi Code Bench 3.0, long tasks (internal) | 55.95% | 61.2% | — |
| Hallucinated claims, Kimi Hallu Bench Hard (internal, lower is better) | 23.4% | 8.3% | — |
Two of the three are Moonshot's own internal tests, so treat them as direction rather than proof. The public one, Terminal-Bench, puts an open-weight model about a point behind Opus 5.5 and seven ahead of GPT-6 Astra on agent work in a terminal. Moonshot also reports gains on its internal finance tasks: financial modelling up 7.2 points over K3 and financial research up 11.6.
What changed for developers
- Thinking is always on.
reasoning_efforttakes low, high or max, and the default is max. A request that tries to switch thinking off is rejected, so a client written for K3 with thinking disabled needs a change. - Same 1M context. 1,048,576 tokens, image and video input, tool calls and strict JSON-schema output.
- Weights by November 4. Moonshot calls it a 3T-class open-weight model and says the full weights will be published by November 4, 2026. They are not on Hugging Face yet, and no license has been named. Moonshot says it took cybersecurity data out of training for the open release.
- Zero data retention is available to API customers.
| Per 1M tokens | Input | Cached input | Output |
|---|---|---|---|
| kimi-k3.1-pro | $2.00 | $0.20 | $10.00 |
| kimi-k3.1-pro-highspeed | $4.00 | $0.20 | $20.00 |
| kimi-k3, for comparison | $3.00 | $0.30 | $15.00 |
The API model unlocks after a top-up of at least $1. In Kimi Code, K3.1 Pro is on the Plus plan and above; the 1M context and Highspeed need Pro or above, and Highspeed spends about twice the quota.
What it means for you
- On K3 in the API? Switch and save. The same work costs a third less per token on K3.1 Pro. Run a sample of real tasks first, because the always-on thinking makes answers slower and uses more output tokens.
- Using Kimi Code on Plus? Type
/modeland pickk3.1-pro. Leave Highspeed for when you are waiting on it, since it burns quota twice as fast. - Still on K2? Moonshot's pricing page says K2 models close to new users on October 31 and retire by the end of December 2026. Plan the move now.
- Want to self-host? Wait for the weights and the license on November 4 before you plan hardware for a model this size.
Our Kimi review has the plans and the updated API prices. For prompts that keep a coding agent working in small, checkable steps, which suits a model that always thinks first, see the Claude Code prompts.
Try it
Answer the question below. Then go back over your answer and list every factual claim in it as a numbered list. For each claim, say: - how sure you are (high / medium / low), - whether it is something you know or something you inferred, - what I should check to confirm it. Mark any number, date, name or quote you are not certain of with [CHECK]. Do not soften the answer to avoid the list. Question: [your question]
You are working in my repository on a task that will take many steps. Before you change anything: 1) restate the goal in two sentences, 2) list the files you expect to touch and why, 3) split the work into steps that each end with something I can run or test, 4) name the step most likely to go wrong and how you'll know. Then do step 1 only, run the check, and stop with a short report. Wait for "next" before each further step. Task: [what you want built or fixed]
Sources
- Kimi, Kimi K3.1 Pro launch post, October 23, 2026
- Kimi API Platform, K3.1 Pro quickstart and chat model pricing, checked October 23, 2026
- Kimi Code docs, models and membership, checked October 23, 2026