Moonshot AI has released Kimi K3.1 Pro, an open-weight model that serves as a post-trained version of the 2.8-trillion-parameter Kimi K3 model. The model achieved a score of 65.7% on Terminal-Bench 4.0, outperforming GPT-6 Astra, which scored 58.2% on the same benchmark.
Claude Opus 5.5 recorded a slightly higher score of 66.4% on Terminal-Bench 4.0. K3.1 Pro leads the SWE-Marathon with a score of 51.7 and achieves 70.0 on Humanity’s Last Exam Full when using tools.
Additional performance metrics include a 92.7% score on BenchCAD with tools, a 58.6 score on Frontier Finance, and a 37.0 in a simulated business operations test. The model also scored 61.2% on Kimi Code Bench 3.0, 57.6 in financial modeling, 70.4 in financial research, 38.1 on Terminal-Bench Science 0.1, 31.5 on ProgramBench, and 39.4 on HLE-Diamond.
Moonshot AI reduced the claim-level hallucination rate to 8.3% on the Kimi Hallu Bench Hard test and solved 81.0% of tasks on an internal xdev-bench-hard test. Full model weights are scheduled for release by 4 November 2026, following the removal of cyber-related data from mid-training and reinforcement learning for the open-weight version.
Pricing for K3.1 Pro is set at $2 per million input tokens and $10 per million output tokens, while cached input costs $0.20 per million tokens. A Highspeed variant of Kimi K3.1 Pro runs two to three times faster than the standard model and costs $4 per million input tokens and $20 per million output tokens.
Users can access the model via Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, with Kimi Enterprise administrators able to enable it for their organisations.
