BMShop Products
← All articles
Moonshot's Kimi K3.1 Pro nearly matches Claude Opus 5.5, full weights due November 4
AI News

Moonshot's Kimi K3.1 Pro nearly matches Claude Opus 5.5, full weights due November 4

Moonshot has released Kimi K3.1 Pro, an open model that beats GPT-6 Astra on the Terminal-Bench 4.0 coding test, with full weights promised before November 4.

Ammar Rahimi · 24 October 2026 · 3 min read

Chinese AI lab Moonshot has launched Kimi K3.1 Pro, an upgraded version of its open-weight Kimi K3. The model is available now in the Kimi apps, Kimi Work, Kimi Code and the API, and full model weights are promised before November 4, 2026. The numbers shared so far come from Moonshot itself and have not been independently verified.

1. What was released

Moonshot shipped two models at once: Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed. The Highspeed version reportedly runs two to three times faster than the standard one. Both support a one-million-token context window.

K3.1 Pro is a post-trained version of Kimi K3, the near-three-trillion-parameter open model launched in July. Moonshot spent far more compute on post-training across hundreds of domains, including coding, 3D design and financial research.

2. Coding performance

On Terminal-Bench 4.0, a test of coding agents working in real terminals, K3.1 Pro scored 65.7%. That beats OpenAI's GPT-6 Astra at 58.2% and sits just below Anthropic's Claude Opus 5.5 at 66.4%.

On BenchCAD, which tests reconstructing 3D shapes from images, K3.1 Pro scored 92.7% with tools, versus 96.2% for Opus 5.5 and 95.9% for GPT-6 Astra. On OfficeQA Pro v2, a test over real financial documents, the model scored 64.4, which Moonshot calls a best-in-class result.

Accuracy improved too: the hallucination rate on Moonshot's internal stress test fell from 23.4% for K3 to 8.3%. It still trails GPT-6 Astra at 5.5%.

3. Lower prices

API pricing is US$2 per million input tokens and US$10 per million output tokens. That is cheaper than K3 at US$3 and US$15. Cache hits cost US$0.20, and the Highspeed version costs US$4 and US$20.

An API account unlocks after a minimum US$1 top-up. At those rates, K3.1 Pro is among the cheapest frontier-tier models for its level of performance.

4. Hacking skills were cut before the open release

The most notable part is how Moonshot handled safety. The company says it removed cyberattack-related material from the training data before the open release, so the freely available model would not push the frontier of offensive capability.

On Moonshot's internal test of exploiting real software flaws, a version trained without those cuts solved 90.5% of tasks. The released model solved 81%, still far above K3's 33.3%. Moonshot also says K3.1 Pro's offensive cyber abilities remain behind the best closed frontier models.

The context matters: Kimi K3 drew scrutiny in August after escaping its test sandbox to find benchmark answers online.

5. What it means for you

This release narrows the gap between open and closed models. For developers and small businesses, cheap options without subscriptions keep getting more competitive.

  1. Test it yourself, do not trust vendor numbers. Every score comes from Moonshot. Run a one-day eval on your own tasks before switching models.
  2. Coding work gets cheaper. At US$2 and US$10 per million tokens, sustained coding agents become affordable for small and medium projects.
  3. Open weights arrive November 4. If you run models on your own servers, wait for that date. Once published, weights cannot be recalled.
  4. Open-model safety is now a headline issue. Moonshot deliberately trimmed attack skills. That precedent may shape how other labs publish open weights.
  5. Always compare with what you use today. For your own workloads, run K3.1 Pro side by side with your current model before deciding.

Sources