Moonshot AI released Kimi K3.1 Pro on October 23, 2026, a heavily retrained version of July's Kimi K3 rather than a new architecture.

In independent tests by Artificial Analysis, the model scores 55 on the Intelligence Index and ranks fourth overall, up from 44 for its predecessor — an 11-point jump.

According to heise online, that would make K3.1 Pro the strongest open-weight model in the ranking if its weights ship as planned, sitting behind Anthropic's Claude Opus 5.5 at 58 and Claude Sonnet 5.5 and Muse Wonder Full at 56 each.

The model is positioned against Claude Opus 5.5 and OpenAI's GPT-6 Astra in coding, 3D modeling and financial analysis.

The striking part is where the gains came from.

As Basic Tutorials notes, Moonshot itself says the model is a retraining, not a major technical leap — the Mixture-of-Experts architecture stays at 2.8 trillion total parameters with 104 billion active per token, and the roughly one-million-token context window is unchanged.

The progress stems mainly from more extensive post-training with reinforcement learning across hundreds of task areas.

Basic Tutorials frames it as proof of how much performance can be squeezed from an existing foundation without billion-dollar investments in a larger base model — and Moonshot cut API prices by about a third at the same time.

The agentic numbers are where the model stands out. Per heise online, K3.1 Pro's Terminal-Bench 4.0 success rate rises from 13% to 67% — the best result in Artificial Analysis's tests — while AutomationBench-AA rule-compliant task completion climbs from 58% to 73%.

In GDPval-AA, which tests professional tasks, it scores 1,846 Elo, ranking third behind Claude Opus 5.5 (1,866) and Muse Wonder Full (1,857). Gains in pure knowledge are smaller: Humanity's Last Exam rises only from 47% to 52%, and the AA-Omniscience hallucination rate falls from 53% to 44%.

Basic Tutorials adds Moonshot's own Terminal Bench figure of 65.7%.

Moonshot charges $2 per million input tokens and $10 per million output tokens, one-third less than Kimi K3, with cache hits at 20 cents per million; the Highspeed version, claimed to be two to three times faster, costs $4 and $20. heise online reports the model generates about 74 output tokens per second versus 42 for K3 and 97 for Claude Opus 5.5, with Highspeed reaching about 129.

But heise also notes the catch: K3.1 Pro burned an average of 55,000 output tokens per Intelligence Index task, 14% more than K3, so a task costs $2.11 versus $2.00 — still far below Claude Opus 5.5 at $5.98 per task.

The open-weight release carries the debate's sharpest edge. Moonshot says it removed cyber-related data from all training stages of the open model while keeping general-purpose coding tasks, warning that without this, downstream cyber capability would increase substantially.

National Technology News reports that on the xdev-bench-hard vulnerability benchmark, the release model solved nearly 10% fewer problems than before the removal, scoring below GPT-5.6 Sol and DeepSeek's V4.1-Flash — and that Moonshot says this lower score is what made it comfortable releasing the weights.

Meanwhile, shattered.io's coverage of the weapon-versus-armour debate around DeepSeek, z.ai and Alibaba notes researchers say removing safeguards does not necessarily increase real-world harm.

The complete model weights are scheduled for release by November 4, 2026, according to heise online, National Technology News and Basic Tutorials — the moment K3.1 Pro becomes downloadable, modifiable and deployable on private hardware.

Until then, the model is live on kimi.com, the mobile apps, Kimi Work, Kimi Code and Moonshot's API under kimi-k3.1-pro and kimi-k3.1-pro-highspeed, though Basic Tutorials notes Kimi Enterprise access is disabled by default at launch.

Basic Tutorials cautions that Moonshot's own benchmark figures will only be validated once independent labs complete their tests, and anyone planning local deployment will need substantial hardware. If the ranking holds, an open-weight file will soon sit one step from the frontier.