Moonshot's Kimi K3.1 Pro ranks fourth on a top AI index ahead of its open-weights release
Moonshot AI's new Kimi K3.1 Pro scores 55 on the Artificial Analysis Intelligence Index, fourth overall, with token prices cut by a third. Full weights are due by November 4.
Moonshot AI has released Kimi K3.1 Pro, a reasoning model that the Beijing company describes as its most capable open-weight system so far, and an independent benchmark puts it fourth among all models it tests. The full weights have not been published yet. Moonshot says they will be released by November 4, and when that happens the model is set to become the highest-scoring open-weights model on that leaderboard.
The launch matters to anyone building on AI models because it narrows the distance between models you can download and run yourself and the closed systems sold only through an API. According to Artificial Analysis, the gap between the best open-weights model and the best proprietary one on its Intelligence Index would shrink from 12 points to 3 once K3.1 Pro's weights are out.
What Moonshot shipped
K3.1 Pro is the first model in a new K3.1 series. It supports a context window of 1 million tokens, which lets it hold a large codebase or a stack of financial filings in a single request. Moonshot calls it a "3T-class" model; Artificial Analysis lists it at 2.8 trillion total parameters with 104 billion active, meaning only a fraction of the network does the work for any given token.
Moonshot says K3.1 Pro is a post-trained version of its earlier K3 model, built with substantially more post-training compute. The company describes running reinforcement learning jointly across hundreds of task domains, and says it reworked its training framework, long-context inference and sandbox infrastructure so that added compute produced steady gains. A faster variant, K3.1 Pro Highspeed, runs at two to three times the speed of the standard model, according to the company.
Both versions are available now on Kimi.ai, the Kimi apps, the Kimi Work desktop app, the Kimi Code coding tool and the Kimi API. For enterprise customers, access is switched off by default and has to be enabled by an administrator.
How it scores
On the Artificial Analysis Intelligence Index, K3.1 Pro scores 55, up 11 points from K3's 44. It trails only Anthropic's Claude Opus 5.5 at 58, and Claude Sonnet 5.5 and Meta's Muse Wonder Full, both at 56. The largest gains came on agent-style work. Artificial Analysis measured 67% on Terminal-Bench 4.0, a test of coding agents working in a constrained terminal, which it says is the highest score of any model it has run. On GDPval-AA, which grades economically useful tasks, the model placed third with 1,846 Elo, up 312 points from K3.
Moonshot's own figures are close to those numbers. The company reports 65.7% on Terminal-Bench 4.0, ahead of OpenAI's GPT-6 Astra at 58.2% and just behind Claude Opus 5.5 at 66.4%. On BenchCAD, which asks a model to rebuild 3D objects from multi-view images by writing CAD code, Moonshot reports 92.7% with tools, short of Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%.
Outside agent tasks the improvements are smaller. Artificial Analysis says Humanity's Last Exam rose from 47% to 52% and SciCode from 59% to 61%. Its hallucination rate for the model fell from 53% to 44%. Moonshot, which is pitching the model at financial research, reports that on an internal stress test of factual accuracy the claim-level hallucination rate dropped from 23.4% for K3 to 8.3%, and it notes that the test is built to provoke errors and does not reflect everyday error rates.
Cheaper tokens, same cost per task
Moonshot cut prices with the release. Artificial Analysis lists first-party pricing of $2 per million input tokens, $10 per million output tokens and $0.20 per million cached input tokens, 33% lower than K3 on all three. The model also thinks longer, though: it used about 55,000 output tokens per Intelligence Index task, 14% more than K3, so Artificial Analysis puts its cost per task at $1.98, roughly the same as its predecessor.
A deliberate limit on cyber capability
The release notes spend unusual space on security. Moonshot says it found cyber-related material in its general training data and reinforcement-learning tasks, and that for the open-weight release it removed that data at both stages. On its in-house exploit benchmark, xdev-bench-hard, a checkpoint trained without those removals solved 90.5% of problems, while the released K3.1 Pro solved 81.0%. The company says the model's offensive cyber capability remains behind frontier proprietary models and that it does not intend to release open weights that push that frontier.
The business behind the model
Moonshot has been raising money at a rapid pace. According to Caixin, the company closed a Series F in July at a $35 billion post-money valuation, led by China's national AI investment fund and other state-backed investors, and in August began raising a pre-IPO round at a $50 billion pre-money valuation. Caixin also reports that Amazon made K3 available through Amazon Bedrock and its Kiro coding assistant in September.
Chinese labs have used open weights to compete with better-funded U.S. rivals since DeepSeek's early 2025 breakthrough showed how far a lower-cost approach could go, a shift we covered at the time in our look at DeepSeek's effect on the industry. Open research groups in the U.S. have pushed in the same direction, including Ai2's open AI infrastructure for science.
The next milestone is the weights release, which Moonshot has committed to complete by November 4, 2026.
