From the source
Kimi released Kimi K3.1 Pro, a 3-trillion-parameter model it calls its most capable, alongside a faster Highspeed variant and a 3D Make feature on kimi.com and the Kimi App.
- The model scores 55 on Artificial Analysis's global intelligence index, fourth worldwide and 11 points above K3, and Kimi says it costs about 80% less than GPT-6 Astra.
- K3.1 Pro supports a million-token context window and targets long-horizon coding, knowledge work and complex reasoning, including 3D creation and financial research.
- It scores 65.7% on Terminal-Bench 4.0, 61.2% on a long-horizon coding test and 92.7% on BenchCAD, and Kimi says it tops Office QA Pro v2 and SpreadsheetBench 2.
- In tests it built a playable Godot game after 23 rounds of self-testing and a Blender keyboard model, and passed a simulated driving exam with 100 points in 151.03 seconds.
- The API is live at 1.3 yuan per million input tokens with cache hits, 13 yuan without, and 68 yuan for output; full weights are promised within two weeks.