Moonshot AI has released Kimi K3.1 Pro, the first model in a new K3.1 series, three months after the original Kimi K3.
It went live on October 23 (October 24 in China) through the Kimi API. Moonshot is pitching it as a sharper agentic and coding model rather than a new architecture: coverage describes it as a post-trained version of K3, with the same 2.8 trillion total parameters (about 104 billion active per token) and a 1-million-token context window.
The benchmark picture
The numbers come from two places, and they are worth keeping apart: Moonshot's own launch table, and independent testing by Artificial Analysis.
Moonshot's reported results
| Benchmark | Kimi K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|
| Terminal-Bench 4.0 | 65.7% | 66.4% | 58.2% |
| Terminal-Bench Science 0.1 | 38.1 | — | 68.1 |
Moonshot also reports 61.2% on its internal Kimi Code Bench 3.0 Long-horizon test, up from K3’s 55.95%. That benchmark is Moonshot’s own, so there is no outside comparison for it.
Two cautions apply. Moonshot says it ran the Terminal-Bench comparison with the Claude Code harness, and that kind of scaffolding can move agentic scores by many points, so this is not a single controlled experiment. And the Science result is a clear weak spot: K3.1 Pro trails GPT-6 Astra by 30 points in Moonshot’s own chart.
Independent results
Artificial Analysis places Kimi K3.1 Pro 4th on its Intelligence Index with a score of 55, an 11-point gain over K3, driven mainly by agentic performance. It says the model will be the leading open-weights model once Moonshot publishes the weights. Coverage of the same testing reports a Terminal-Bench score of 67%, the highest Artificial Analysis has measured, which would put K3.1 Pro ahead of Opus 5.5 on that run. That conflicts with Moonshot’s ordering, a reminder that the two runs used different setups.
On hallucinations, Moonshot reports a large drop on its internal stress test, while Artificial Analysis measured a smaller improvement. Published figures also differ between Moonshot’s launch post and its API quickstart, so treat any single number here as provisional.
A model trained to be less dangerous at hacking
The most unusual part of the release is what Moonshot removed. According to reporting from The Next Web and others, Moonshot filtered cyber-offense material out of the training data before the planned open-weights release. Its in-house exploit test fell from 90.5% on the unfiltered checkpoint to 81% on the released model.
The filtering did not leave the model weak at security work: Artificial Analysis’s Cyber Index ranks K3.1 Pro first among publicly available models at 64%. For security teams, the takeaway is that “safer” here means reduced, not eliminated.
Price and access
K3.1 Pro is available through the Kimi API as kimi-k3.1-pro, and a faster kimi-k3.1-pro-highspeed variant is also listed. According to launch coverage and the API platform’s documentation:
- Standard: $2 per million input tokens, $10 per million output tokens, with cached input at $0.20.
- High-speed: $4 input and $20 output per million tokens.
- Context: The full 1M window is billed at a flat per-token rate with no long-context surcharge.
- Access: The quickstart says the model unlocks after a top-up of at least $1.
That is cheaper per token than K3’s launch pricing, which matters for teams running long agent loops where output tokens add up.
What is still unconfirmed
- Open weights. Moonshot says weights will be published by November 4, 2026. As of launch there was no K3.1 Pro repository on Hugging Face.
- License. Moonshot had not said whether K3.1 Pro will use the same license as K3. Check the terms before building a product on it.
- Independent replication. Most coding and agent figures are vendor-reported. Expect more third-party runs once developers can test it and the weights are public.
Official teaser
Moonshot’s official Kimi account teased the release the day before launch with a 24-second retro-style clip. It shows a Kimi Code terminal reading “Model: Kimi K3.1” while it writes a procedural 3D scene in Three.js. The post carried no caption. We could not confirm a public embeddable recording of the launch itself, so the clip is not embedded here; it is available on the Kimi account on X.
Why it matters
For professionals who rely on AI coding and agent tools, K3.1 Pro is another sign that the price gap between frontier and near-frontier models is narrowing. A model that sits within a point or two of Claude Opus 5.5 on one coding benchmark, at a fraction of the token price, will get tested by teams that currently pay premium rates. The open-weights promise adds a second reason to watch: if the weights arrive on schedule and the independent results hold, the model could be run on your own infrastructure.
The caveat is the usual one. Benchmarks reflect harnesses, prompts and settings, and the vendor’s own table is the least independent evidence in the set. Run it on your own tasks before changing a workflow.