Kimi K3.1 Pro: Near-Frontier Scores, Open Weights Due Nov 4

Moonshot AI’s Kimi K3.1 Pro scores 55 on the Artificial Analysis Intelligence Index, 3 points behind Claude Opus 5.5, and Moonshot says the full weights will be public by November 4. Simon Willison calls it the first time a soon-to-be open-weight model has scored this highly against the frontier. For teams running agents on closed APIs, it means a model they could host themselves may soon be close enough to test as a real alternative.

What Happened

Kimi K3.1 Pro landed on the Kimi API on October 23, along with a Highspeed version that Moonshot says runs two to three times faster. Willison notes it costs twice as much, except for cached input. According to Moonshot’s technical blog, both are also live in Kimi.ai, Kimi Work and Kimi Code, and the weights will follow by November 4, 2026.

The Decoder reports that K3.1 Pro is a post-trained version of K3 with much more post-training compute. It keeps K3’s size of 2.8 trillion total and 104 billion active parameters, with a one-million-token context window and text and image input. The 55 score is 11 points above K3.

Why It Matters

The gains are concentrated in agentic work. On Terminal-Bench 4.0, Moonshot reports 65.7%, against 66.4% for Claude Opus 5.5, 58.2% for GPT-6 Astra and 13% for K3. Artificial Analysis puts it at 1,846 Elo on GDPval-AA, 312 points above K3. Moonshot also reports claim-level hallucinations on its internal Hallu Bench Hard falling from 23.4% to 8.3%, while noting it’s a stress test, not a general error rate. Most of these numbers come from Moonshot, so they need independent checking.

Prices fell by a third: $2 per million input tokens, $10 per million output tokens and $0.20 for cache hits, down from $3, $15 and $0.30. Willison notes that’s half the price of Claude Opus 5.5. The Decoder adds a caveat: K3.1 Pro generates 14% more output tokens, so a full Intelligence Index task costs $1.98, almost the same as K3’s $2.00.

Moonshot also limited one capability on purpose. It says vulnerability research skill kept rising with post-training, so it removed cyber-related data from training for the open-weight release. On its internal xdev-bench-hard, the uncleaned checkpoint solves 90.5% of problems and the released model 81.0%. Moonshot says it doesn’t want open weights to extend the frontier of offensive cyber capability.

What This Means for Agent Builders

Two practical deadlines sit alongside the launch. Kimi’s API pricing page says K2 series models will close to new users on October 31, 2026, and Moonshot plans to retire all of them by the end of December. And when the weights arrive, the license matters as much as the benchmarks: Stack Archive covered how Kimi K3’s commercial license got tighter at its open release.

Source: Simon Willison's Weblog

Content created with AI assistance and reviewed for accuracy.