Kimi K3.1 Pro Closes the Open-Weight Gap

Moonshot AI released Kimi K3.1 Pro on October 23. It is a reasoning model with a 1M-token context window, available now through the Kimi API, and Moonshot says the full weights ship by November 4.

That date is the story. If the weights land as promised, the best open model on the market will sit within a few points of the best closed ones on agentic coding.

Image: Moonshot AI, Introducing Kimi K3.1 Pro, October 23, 2026.

What shipped

K3.1 Pro is a post-trained version of Kimi K3. Moonshot says the gain comes from much more reinforcement learning compute, not a new base model. A Highspeed variant runs two to three times faster, per Moonshot.

The API model IDs are kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Thinking is always on. Reasoning effort takes low, high, or max, and max is the default. Temperature and top_p are fixed, so omit them. Output defaults to 131,072 tokens and can be raised to the full 1,048,576. The model unlocks after a top-up of at least $1. Vision input does not accept public image URLs; send base64 instead.

Vercel listed both models on AI Gateway the same day. That means teams already routing Claude Code or Codex through the gateway can point an agent at moonshotai/kimi-k3.1-pro without a new vendor account. Kimi Enterprise admins must switch the model on; it is off by default.

The numbers, and who reported them

Moonshot's own figures: 65.7% on Terminal-Bench 4.0, run with a Claude Code harness. It lists GPT-6 Astra at 58.2% and Claude Opus 5.5 at 66.4%. On BenchCAD with tools, it reports 92.7% against 96.2% for Opus 5.5. These are vendor numbers and have not been independently reproduced.

Artificial Analysis ran its own suite. It scores K3.1 Pro at 55 on its Intelligence Index, fourth overall and up from 44 for Kimi K3. Only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56), and Muse Wonder Full (56) score higher. In its own Terminal-Bench 4.0 run it measured 67%, the highest of any model it has tested. The two Terminal-Bench figures come from different runs, so do not compare them directly.

Price per token is down; price per task is flat. Artificial Analysis lists $2 input, $10 output, and $0.20 cache hits per million tokens, 33% below K3. But K3.1 Pro used 14% more output tokens per task, so its measured cost per Intelligence Index task, $1.98, is roughly the same as K3. It also reports 2.8T total parameters with 104B active.

The cyber trade-off

Moonshot says it removed cyber-related data from mid-training and from its RL environments before the open release. On its internal exploit benchmark, the unfiltered checkpoint solved 90.5% of tasks at pass@3. The released model solved 81.0%. Moonshot says that still trails frontier closed models. It is an unusual disclosure, and it is entirely self-reported.

Hallucination claims need the same caution. Moonshot's blog cites a drop in claim-level hallucination from 23.4% to 8.3% on an internal stress test. Its API quickstart cites a different figure for what reads as the same test. Treat both as directional.

What to do this week

Run it on your own eval set, not the launch table. Test reasoning_effort high against max, because always-on thinking is where the output-token bill comes from. If you use AI Gateway, swap the model ID in one agent profile and compare pass rates and cost per merged change.

Do not plan self-hosting yet. As of launch, Artificial Analysis notes the weights are unpublished, and the license terms are not stated. Revisit after November 4. For API use, Moonshot says zero data retention is available to API customers; confirm it is on before you send proprietary code.

Sources

https://www.kimi.ai/blog/kimi-k3-1-pro

https://platform.kimi.ai/docs/guide/kimi-k3-1-pro-quickstart

https://artificialanalysis.ai/articles/kimi-k3-1-pro-launch-benchmarks

https://vercel.com/changelog/kimi-k3-1-pro-and-kimi-k3-1-pro-highspeed-are-available-now-in-ai-gateway

https://thenextweb.com/news/moonshot-kimi-k3-1-pro-open-weight-cyber