Moonshot AI releases Kimi K3.1 Pro, now fourth on the Artificial Analysis index

Moonshot AI has released Kimi K3.1 Pro, an upgrade to its flagship Kimi K3 from July 2026. The company says the gain comes from a much larger post-training run on the same base model rather than a new architecture. A faster variant, K3.1 Pro Highspeed, arrived with it. It is described as two to three times quicker at the same capability, so users can choose speed without giving up quality. Both are live on Kimi.ai, Kimi Work, Kimi Code and the API. Reports are dated 23 October in US time, though Unite.AI dates the launch to 24 October. K3.1 Pro keeps K3's dimensions, so the change is in how the model was trained, not in its size.

The headline result is on the Artificial Analysis Intelligence Index, where K3.1 Pro scores 55, eleven points above K3. That puts it fourth, behind Claude Opus 5.5 at 58 and Claude Sonnet 5.5 and Muse Wonder Full at 56 each. Coding is where the jump is biggest: on Terminal-Bench 4.0 it reaches 65.7%, against 13% for K3, 58.2% for GPT-6 Astra and 66.4% for Opus 5.5. It trails Opus 5.5 and Astra on BenchCAD, at 92.7% with tools, and leads on OfficeQA Pro v2. Moonshot's own hard hallucination test shows errors falling from 23.4% to 8.3%, while Astra sits at 5.5%. On cost per task, the Artificial Analysis comparison places K3.1 Pro on the efficiency frontier at $1.98 against $2.00 for K3, so the extra capability comes at no higher cost per task.

The model takes text and image input with a one-million-token context window. Kimi Code's documentation also lists video input and a thinking-only mode with low, high and max effort. The Decoder also points to a new 3D Make feature on Kimi.com is aimed at Unity, Unreal, Godot, Blender and CAD scripts. API prices fall by a third, to $2 per million input tokens and $10 per million output tokens, with cached input at $0.20 per million. That is a 33% reduction on the previous API rate card.

Outlets disagree on how big the model is. SiliconANGLE reports 3 trillion parameters across 896 experts with 16 active, The Decoder gives 2.8 trillion in total with 104 billion active, and Kimi Code's docs call it a 3T-parameter flagship. SiliconANGLE also misspells the name as "Kim K3.1 Pro" in places.

Moonshot says it stripped cyber-related data from both training phases of the model it plans to open. Its reasoning is that no open model should push the offensive frontier, but open models should still help defenders audit old code. On a hard vulnerability-finding test the model scores 90.5% before the data was cleaned and 81.0% after, compared with 33.3% for K3 and 85.7% for DeepSeek-V4.1-Flash. All three outlets report the cyber-data removal as a deliberate choice rather than a limitation.

Moonshot has promised the open weights by 4 November 2026, which all three outlets note as the next milestone to watch. If they ship, the reporting notes this would make K3.1 Pro the top-scoring open-weights model on the Artificial Analysis index.