Moonshot AI, the Beijing company behind the Kimi assistant, released Kimi K3.1 Pro in late October and says it scored 65.7% on Terminal-Bench 4.0. By Moonshot's count that is 0.7 points behind Claude Opus 5.5 and well ahead of GPT-6 Astra. Moonshot also says the full model weights will be published by 4 November, which would make it one of the strongest openly available models so far.
What Kimi K3.1 Pro is
According to Moonshot's blog and developer docs, K3.1 Pro is the new top model in the Kimi series. The headline details are:
- A context window of about 1 million tokens, which is enough for a large codebase or a long stack of documents.
- Availability now in the Kimi apps and through the Kimi API, with a faster companion called K3.1 Pro Highspeed.
- Full model weights promised by 4 November 2026.
- API access that switches on after a minimum top-up of $1, according to Moonshot's quickstart guide.
The benchmark numbers, and who produced them
All of the figures below come from Moonshot. On Terminal-Bench 4.0, a test of coding agents working in a terminal, Moonshot reports 65.7% for K3.1 Pro, 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra. Moonshot says it ran its own model with the Claude Code harness and took the other scores from published results.
The model does not win everywhere. On BenchCAD with tools, Moonshot's chart shows 92.7 for K3.1 Pro against 96.2 for Opus 5.5 and 95.9 for GPT-6 Astra. Trade publication VKTR, working from Moonshot's charts, also notes much lower scores on a science-focused terminal test. VKTR points out that several of the evaluations are internal and have not been independently verified, and that is the right way to read them.
Fewer made-up claims
The biggest improvement Moonshot reports is in hallucination. On its own adversarial test with tools, the share of claims that were wrong fell from 23.4% on K3 to 8.3% on K3.1 Pro. That is a large drop, though it is again an internal benchmark designed by the company that made the model.
The skill Moonshot held back on purpose
One detail stands out. Moonshot says it removed cyber-related training data before the open-weight release. On its internal hacking benchmark, a version without those measures scored 90.5%, and the released model scored 81.0%. The company says it does not want open weights to push the frontier of cyber capability, and that K3.1 Pro stays behind proprietary models that have been public for at least three months. It is a notable choice for a lab that publishes its weights.
The Kimi K2 cutoff
If you build on older Kimi models, Moonshot's model list and pricing page say that from 31 October 2026, K2 series models are no longer available to newly registered users, while existing users are not affected for now. Moonshot plans to retire all K2 series models by the end of December 2026 and points developers to kimi-k3.1-pro as the replacement. Which exact model IDs count as the K2 series is not spelled out in the notice itself, so check the model list for the ones you use.
What to do this week
Pick three real tasks you gave Claude or ChatGPT last week, run the same prompts through Kimi K3.1 Pro, and score the answers side by side. Before you paste anything sensitive, read Moonshot's data terms, the same way you would for any new provider. If you run anything on K2, put the end of December in your calendar and plan the switch now. For a broader view of how the main assistants compare, see our comparison page, and for background on the Chinese AI market, read our piece on the China AI stock slide.