Moonshot AI releases Kimi K3.1 Pro, scoring 55 on Artificial Analysis Intelligence Index, with 33% lower API prices and open weights planned by November 4

Moonshot AI released Kimi K3.1 Pro, a further post-trained version of Kimi K3 with the same size (2.8T total, 104B active parameters, 1M context, text and image input). It also released a 2-3x faster Highspeed variant. Both are available now through Kimi.ai, Kimi Work, Kimi Code and the Kimi API. API prices fell about 33% to $2 input and $10 output per million tokens. Artificial Analysis, which is independent of Moonshot, scores it 55 on its Intelligence Index, up 11 from K3 and fourth behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56). Open weights are promised by about 4 November 2026 and have not been released.

Teams that buy model access get a near-frontier option at a lower price, and it is available today. That gives them some leverage when negotiating with proprietary vendors. The larger effect depends on the open weights. If they arrive, organisations that need to self-host could use a model close to the proprietary leaders, and the open-versus-proprietary gap would shrink from 12 index points to 3. Until then this is a hosted API from a Chinese vendor, so data residency, compliance and procurement rules may rule it out for some buyers. The weights also have a cost. Moonshot removed cyber-related data from training for the open release, and its own internal cyber benchmark fell from 90.5% to 81.0%. The released model may therefore differ from the hosted one.

The headline "closing in on Claude Opus 5.5" rests on one index, where three points is a modest gap. The Terminal-Bench jump (13% to 65.7%) and the hallucination drop (23.4% to 8.3%) are Moonshot's own figures, and the hallucination result comes from an internal benchmark nobody else can check. The "highest-scoring open-weights model" claim is conditional on a release that is only promised.

1. 01Whether the weights actually appear by 4 November 2026, with what licence, and whether the cyber-filtered version matches the hosted model's score
2. 02Independent replication of Terminal-Bench 4.0 and hallucination results by third parties, plus any change to Artificial Analysis's score once the open weights are tested
3. 03Competitor price moves and real usage: whether Anthropic or other vendors cut prices, and whether Kimi API adoption or third-party hosting of the weights grows in the weeks after release