The gains are concentrated in agentic work, meaning tasks where the model acts and uses tools rather than just answering. On the GDPval-AA benchmark it reaches 1,846 Elo, 312 points above K3 and just behind Claude Opus 5.5. Moonshot reports 65.7 percent on Terminal-Bench 4.0, up from 13 percent for K3 and only slightly behind Claude Opus 5.5's 66.4 percent. A faster variant, K3.1 Pro Highspeed, runs two to three times quicker according to the company, and both models are available now through Kimi.ai, Kimi Work, Kimi Code and the Kimi API.

Moonshot also reports sharply lower hallucination rates, with claim-level errors on its internal Kimi Hallu Bench Hard falling from 23.4 percent to 8.3 percent, though it stresses this is a stress test rather than a measure of general error rates, and GPT-6 Astra still scores better at 5.5 percent. The model keeps K3's dimensions: 2.8 trillion total parameters, 104 billion active ones, a one-million-token context window and text and image input. Token prices drop by a third, to $2 per million input tokens and $10 per million output tokens.

One caveat concerns the open-weights release. Moonshot says it plans to publish the model weights by 4 November 2026, which would make K3.1 Pro the highest-scoring open-weights model available and shrink the gap to proprietary models from 12 points to 3, assuming nothing else changes by then. The company also says it stripped cyber-related data from the open-weights training, cutting its score on an internal benchmark that reproduces recently disclosed vulnerabilities from 90.5 percent to 81.0 percent, because it believes no open-weights model should push the frontier of offensive cyber capabilities.