The gap narrows
Moonshot AI has unveiled Kimi K3.1 Pro, a further-trained version of the previous Kimi K3. According to The Decoder, the model scores 55 points on the Artificial Analysis Intelligence Index, an 11-point gain over the previous version that places it fourth in the ranking. Ahead of it remain only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56). The most significant figure, however, concerns the outlook: if Moonshot publishes the weights as expected by November 4, 2026, K3.1 Pro would become the highest-scoring open-weights model and the gap between open and proprietary models would fall from 12 to 3 points, provided the ranking does not change in the meantime.
Alongside K3.1 Pro, Moonshot is making available a variant called K3.1 Pro Highspeed, which the company says is two to three times faster. Both models are already accessible via Kimi.ai, Kimi Work, Kimi Code and the Kimi API.
Where Moonshot says it improves
The sharpest gains concern agentic tasks. On the GDPval-AA benchmark, the model reaches 1,846 Elo according to Artificial Analysis, 312 points more than K3, placing it just behind Claude Opus 5.5 (1,866) and Muse Wonder Full (1,857). On Terminal-Bench 4.0, Moonshot reports 65.7 percent, ahead of GPT-6 Astra (58.2 percent) and just below Claude Opus 5.5 (66.4 percent); K3 stopped at 13 percent. Hallucinations also decline: on the internal Kimi Hallu Bench Hard test, which the company describes as a stress test rather than a general measure of errors, the per-claim rate falls from 23.4 to 8.3 percent, while GPT-6 Astra stands at 5.5 percent.
On internal financial benchmarks, K3.1 Pro gains 7.2 points in Financial Modeling and 11.6 in Financial Research; for office tasks, Moonshot names it the best among the models tested on OfficeQA Pro v2. The company also highlights its ability to work directly with engines such as Unity, Unreal Engine and Godot, and on BenchCAD the model reaches 92.7 percent with tool use, close to Claude Opus 5.5 (96.2 percent) and GPT-6 Astra (95.9 percent), versus 72.8 percent without tools.
Pricing, safety and points to verify
K3.1 Pro retains K3's dimensions: 2.8 trillion total parameters and 104 billion active, with a context window of one million tokens and the ability to handle text and images. According to Artificial Analysis, it sits at the top of the Pareto frontier; no model with a comparable number of parameters scores higher. Prices fall by 33 percent: $2 per million input tokens (down from $3), $10 for output (down from $15) and $0.20 for cache hits (down from $0.30); the per-task cost on the Intelligence Index remains almost identical to K3 ($1.98 versus $2.00) because the model generates 14 percent more output tokens.
Moonshot devotes considerable space to cybersecurity: the company says it observed a growth in vulnerability-research capabilities as post-training progressed, because cyber content was present in the mid-training corpus and in reinforcement-learning tasks. For the open-weights release, that data was removed from both training phases: on the internal xdev-bench-hard test, the uncleaned version solved 90.5 percent of tasks, the cleaned version 81.0 percent, versus 85.7 percent for DeepSeek-V4.1-Flash and 33.3 percent for Kimi K3. The results on internal benchmarks, and above all compliance with the expected date for publication of the weights, remain to be independently verified.
