Moonshot AI released Kimi K3.1 Pro on October 23, 2026, with open weights scheduled for November 4, 2026. A podcast discussion following the launch raised concerns about the model's promotion and certain performance gaps, adding friction to what was otherwise a strong benchmark showing.
The model uses a mixture-of-experts architecture with 2.8 trillion total parameters and 104 billion active parameters, and supports a 1 million-token context window. Pricing is $2.00 per million input tokens and $10 per million output tokens, matching GPT-6.1 Sol and half the price of Claude Opus 5.5. A Highspeed variant runs 2-3x faster at twice the price, except for cached input tokens.
Artificial Analysis scored K3.1 Pro at 55, ahead of Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon (all at 53), trailing only Claude Opus 5.5 at 58. Simon Willison noted this marks 'the first time a (soon-to-be) open weight model has scored this highly against the frontier.' On Terminal-Bench 4.0, the model scored 65.7%, 0.7 points behind Claude Opus 5.5 and 7.5 points ahead of GPT-6 Astra at lower average cost per task. Performance is uneven: it leads on OfficeQA Pro v2 (64.4 vs Claude Opus 5.5's 62.2) and Agents' Last Exam (43.2), but trails on Terminal-Bench Science (38.1 vs GPT-6 Astra's 68.1). Post-launch discussion also identified weaker results on ProgramBench as a gap on long-horizon tasks.
The launch was reportedly accompanied by a paid social media campaign that instructed participants not to disclose the promotion. One person shared a screenshot of press instructions stating 'Don't say this is a paid promotion.'
Serving the model requires B300 hardware with one node per model replica, making it substantially more resource-intensive than smaller alternatives. The model's RL training appears optimized for exploitation rather than exploration, which may explain its gaps on open-ended, long-horizon benchmarks.
The most notable pre-release decision was Moonshot's deliberate reduction of exploit-solving capability. After observing exploit skill climb as post-training scaled, Moonshot removed cyber-related data from mid-training and reinforcement learning. On its internal xdev-bench-hard, pass@3 dropped from 90.5% on the unfiltered model to 81.0% on the released version, still far above K3's 33.3%. Moonshot attributes the remaining capability to general coding strength rather than security-specific training. A third-party run of GPT-5.6 Sol solved 100% of the same tasks. Tool-use hallucination on Moonshot's internal stress test fell from 23.4% on K3 to 8.3% on K3.1 Pro, though GPT-6 Astra still leads at 5.5%. Finance hallucinations are down 65% compared to prior Kimi versions. The model also supports full-project coding workflows in Unreal Engine, Unity, Godot, and parametric 3D modeling in Blender and CAD tools. Inference providers are receiving exclusive early access to weights ahead of the public November 4 release.