Moonshot AI's Kimi K3.1 Pro nears Claude Opus 5.5, open weights due by 4 November

Moonshot AI released Kimi K3.1 Pro on 23 October 2026, calling it its most capable 3T-class open-weight model, with a 1-million-token context window [1]. Artificial Analysis ranks it fourth on its Intelligence Index with 55 points [2]. The model is live on Kimi's apps and API now; the full weights are promised by 4 November [1].

What did Moonshot AI release?

Kimi K3.1 Pro is the first model in Moonshot's new K3.1 series and a post-trained version of Kimi K3, built with substantially more post-training compute and large-scale reinforcement learning across hundreds of domains [1]. Alongside it, Moonshot is releasing Kimi K3.1 Pro Highspeed, which the company says runs two to three times faster [1].

Moonshot AI says Kimi K3.1 Pro has a 1-million-token context window and full weights will follow by 4 November 2026. Both models are available now on Kimi.ai, Kimi Work, Kimi Code and the Kimi API [1]. According to Artificial Analysis, the model has 2.8 trillion total and 104 billion active parameters, the same size as Kimi K3, and accepts text and image input [2][4].

API prices are 2 dollars per million input tokens and 10 dollars per million output tokens, 33 percent below Kimi K3 [2]. Simon Willison notes that is half the price of Claude Opus 5.5; The Decoder lists the Highspeed variant at 4 and 20 dollars [3][4].

How does Kimi K3.1 Pro compare with GPT-6 Astra and Claude Opus 5.5?

On Terminal-Bench 4.0, Moonshot reports 65.7 percent for K3.1 Pro, ahead of GPT-6 Astra at 58.2 percent and narrowly behind Claude Opus 5.5 at 66.4 percent [1]. On the BenchCAD 3D-reconstruction test, it scores 92.7 percent with tools, against 96.2 percent for Claude Opus 5.5 and 95.9 percent for GPT-6 Astra, according to Moonshot [1].

Independent testing points the same way. Artificial Analysis scores Kimi K3.1 Pro at 55 on its Intelligence Index, fourth behind Claude Opus 5.5 at 58. Only Claude Sonnet 5.5 and Muse Wonder Full, both at 56, also rank higher, and the score is 11 points above Kimi K3 [2]. In Artificial Analysis' own run, K3.1 Pro reached 67 percent on Terminal-Bench 4.0, the highest score of any model, according to Artificial Analysis [2].

Moonshot also claims gains on its internal tests. On Kimi Hallu Bench Hard, a stress test for factual accuracy, the claim-level hallucination rate fell from 23.4 percent for K3 to 8.3 percent, though Moonshot says the rates do not represent everyday error rates [1].

Why did Moonshot hold back cyber capability?

Moonshot AI says it stripped cyber-related training data so the open weights stay behind frontier cyber capability. The company found cyber content embedded in its general mid-training corpus and RL environment mix, and removed it from both stages for the open-weight release [1].

On its in-house xdev-bench-hard exploit benchmark, a checkpoint trained without these measures solved 90.5 percent of problems, while the final K3.1 Pro reached 81.0 percent [1]. Moonshot says K3.1 Pro's cyber capability remains behind frontier proprietary models that have been publicly available for at least three months [1].

What changes once the weights are out?

Artificial Analysis says K3.1 Pro will be the highest-scoring open-weights model once the weights are published, narrowing the gap between the leading open and proprietary models on its index from 12 points to 3 [2]. Simon Willison calls it the first time a soon-to-be open-weight model has scored this highly against the frontier [3]. For now, the weights are not yet published, Artificial Analysis notes [2].

Our read

Opinion: A three-point gap between the best open model and the best closed one on an independent index would be the story of the year for open weights, but it only counts once the files are actually downloadable. The cyber section is the more interesting precedent: a lab publicly measuring what it removed before release is something the closed labs should be asked to match.

What we could not verify

• Moonshot's own benchmark comparisons with GPT-6 Astra and Claude Opus 5.5; most were run by Moonshot with its chosen harnesses, and some competitor scores are taken from other reports.
• Whether the full weights actually ship by 4 November 2026 and under which licence; Moonshot has not published licence terms in its announcement.
• The Highspeed variant's two-to-three-times speed-up; that is Moonshot's claim and we have not seen independent speed tests of it.
• Results on Moonshot's internal benchmarks, including Kimi Code Bench, the finance suites, Kimi Hallu Bench Hard and xdev-bench-hard, which outsiders cannot reproduce.