What Was Announced?

Moonshot AI has launched Kimi K3.1 Pro, the first model in its new K3.1 series, which the Chinese lab describes as its “most capable 3T-class open-weight model to date.” Announced on October 23, 2026, the model is not a new base model: Moonshot says it is a post-trained version of Kimi K3 put through significantly more reinforcement learning compute across what it calls “hundreds of downstream domains.”

The release arrives amid fierce competition among Chinese labs. Newer models from DeepSeek and Z.ai had recently edged past the original K3 on key leaderboards, creating pressure for Moonshot to iterate quickly. With K3.1 Pro, the company claims to have nearly closed the gap with the world’s top closed models, in some coding tests landing within a single point of Anthropic’s Claude Opus 5.5.

Key Features and Details

Kimi K3.1 Pro keeps the scale of its predecessor: 2.8 trillion total parameters with 104 billion active per token in a mixture-of-experts architecture, and a 1-million-token context window. Alongside it, Moonshot introduced K3.1 Pro Highspeed, a variant it says delivers the same capabilities with 2 to 3 times faster output, consuming roughly twice the quota in Kimi Code.

Every performance number below comes from Moonshot’s own testing and has not been independently verified. On Terminal-Bench 4.0, which tests coding agents on real terminal tasks, Moonshot reports 65.7% for K3.1 Pro, against 66.4% for Claude Opus 5.5 and 58.2% for OpenAI’s GPT-6 Astra. The company says it ran its own model in the Claude Code harness and took rival scores from their published reports and the official leaderboard.

The model also scores 92.7% on BenchCAD with tools (behind Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%, but just ahead of Claude Fable 5.1 at 92.6%), claims a state-of-the-art 64.4 on OfficeQA Pro v2, and cuts the claim-level hallucination rate on Moonshot’s hardest internal factual-accuracy test from 23.4% to 8.3%.

Bar chart of Moonshot's BenchCAD results: Claude Opus 5.5 96.2, GPT-6 Astra 95.9, Kimi K3.1 Pro 92.7, Claude Fable 5.1 92.6
Moonshot’s BenchCAD results for Kimi K3.1 Pro against rival models. Image: Moonshot AI

Independent evaluators are already weighing in. Artificial Analysis, which scores K3.1 Pro at 55 on its Intelligence Index (fourth overall, behind Claude Opus 5.5 at 58, Claude Sonnet 5.5 at 56 and Muse Wonder Full at 56), said on X that the model “will be the leading open-weights model once its weights are published.” On GDPval-AA, an agentic benchmark, K3.1 Pro reached 1,846 Elo, up 312 points over K3 and third behind Opus 5.5 and Muse Wonder Full.

Notably, Moonshot says it deliberately held back the model’s cyber capabilities: after an unredacted version solved 90.5% of internal exploit tasks on Linux kernel, Node.js and Exim, the lab removed cyber-related data from mid-training and reinforcement learning, saying it does not intend to “release open weights that would extend the absolute frontier of cyber capability.” Even so, the model scores 81% on the xdev-bench-hard benchmark.

Bar chart of Moonshot's xdev-bench-hard cyber benchmark: unredacted checkpoint 90.5%, Kimi K3.1 Pro 81.0%
Moonshot’s xdev-bench-hard results showing the cyber-capability gap between the unredacted checkpoint and the released K3.1 Pro. Image: Moonshot AI

Pricing and Availability

Kimi K3.1 Pro is available now on Kimi.ai, Moonshot’s mobile apps, Kimi Work, Kimi Code and via API. Moonshot has cut API prices by 33% across input, output and cache-hit rates compared with K3, while listing the cost per task at roughly $1.98, in line with the earlier model.

The full model weights are a separate milestone: Moonshot promises to release them by November 4, 2026, at which point developers will be able to download and self-host the model. That release is what will let outside testers verify the company’s benchmark claims for themselves.

Why It Matters

Kimi K3.1 Pro is Moonshot’s strongest evidence yet that an open-weight model can keep pace with the best closed systems on agentic coding work. If the weights ship on schedule, it will likely become the highest-scoring open model on the Artificial Analysis Intelligence Index, giving teams that need to keep data in-house a frontier-class option they can host themselves. The caveat remains the usual one: the headline numbers are self-reported, so treat them as directional until independent labs publish results.

Sources