Kimi K3.1 Pro launches: 2.8T parameters, 1M context, prices down 33%, fourth on the global index

Pin the source first. The primary material is a technical blog Moonshot AI published on October 23, 2026 (already October 24 in China), titled Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence at kimi.ai/blog/kimi-k3-1-pro, accompanied by a launch post on Kimi's X account. The Decoder, Unite.AI, Wccftech and OfficeChai followed on October 23. Benchmark figures here lean on Artificial Analysis's independent evaluation; vendor-run numbers are labelled as such, because blending the two turns a company claim into apparent third-party verification.

What the model is. K3.1 Pro is not a new architecture; it is a post-trained version of Kimi K3 with substantially more post-training compute, per Moonshot. The specs are unchanged: 2.8 trillion total parameters, 104 billion active per token in a mixture-of-experts design, a native 1M-token context, and both text and image input. A second variant called K3.1 Pro Highspeed is claimed to be two to three times faster at double the price, $4 input and $20 output per million tokens. Both are available now through Kimi.ai, Kimi Work, Kimi Code and the Kimi API.

What the independent index says. Artificial Analysis puts it at 55 on its Intelligence Index (The Decoder records 55, some outlets 55.1), 11 points above Kimi K3 and fourth overall. The top three are Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56). Artificial Analysis also places it on the Pareto frontier, meaning nothing with a comparable parameter count scores higher. One easily missed fact: the gap between fourth and first is 3 points, and a year ago that gap was in double digits.

The weights have a date, not a release. Moonshot plans to publish the model weights by November 4, 2026. If that holds, K3.1 Pro becomes the highest-scoring open-weights model available and the open-versus-proprietary gap narrows from 12 points to 3. Two preconditions belong in the sentence: that the date holds, and that nothing else ships in the meantime. The Decoder's phrasing is assuming nothing else changes by then, and that qualifier is not optional. Today what you can actually use is the API and Moonshot's own products, so saying open models have caught up is roughly ten days early.

Agentic and coding work improved the most. On Terminal-Bench 4.0 Moonshot reports 65.7 percent, against 13 percent for Kimi K3, 66.4 for Claude Opus 5.5 and 58.2 for GPT-6 Astra. SWE Marathon v1.1 lands at 51.7, which is ahead of Opus 5.5 at 46.3. GDPval-AA reaches 1,846 Elo, 312 points above K3 and behind only Opus 5.5 (1,866) and Muse Wonder Full (1,857). The three numbers point at one thing: this generation's gains sit in tasks that run many steps and require self-correction, not in single-turn answers.

Hallucination rates. On Moonshot's internal Kimi Hallu Bench Hard, claim-level hallucination fell from 23.4 percent on K3 to 8.3 percent. GPT-6 Astra sits at 5.5 percent on the same benchmark, still ahead. Moonshot itself stresses this is a stress test rather than a general error rate, and that caveat should stay. Worth noting in passing: driving down hallucination and driving down tokens per task are two different things, as the next section shows.

Finance, office work and 3D. OfficeQA Pro v2 comes in at 64.4, which Moonshot says leads all tested models, with Claude Opus 5.5 at 62.2 and GPT-6 Astra at 54.4. On internal finance benchmarks it gains 7.2 points on Financial Modeling and 11.6 on Financial Research. 3D is the new emphasis this round: 92.7 percent on BenchCAD with tools and 72.8 without (Opus 5.5 scores 73.0 without tools). It drives Unity, Unreal Engine and Godot directly, and can write and run Blender and CAD scripts without MCP connectors. Kimi.com also added 3D Make for previewing, editing and exporting models.

The cybersecurity section deserves its own read. Moonshot says vulnerability research capability kept climbing as post-training scaled, because cyber-related content was already in the mid-training corpus and related tasks were part of the reinforcement learning environment mix. To keep an open-weights release from pushing the absolute frontier of offensive cyber capability, Moonshot stripped cyber data from both training phases. The cost is measurable: on the internal xdev-bench-hard, which reproduces recently disclosed vulnerabilities in the Linux kernel, Node.js and JavaScriptCore among others, the uncleaned version solves 90.5 percent while cleaned K3.1 Pro solves 81.0. DeepSeek-V4.1-Flash reaches 85.7 on the same test; Kimi K3 managed 33.3. Moonshot's stated position is that no open-weights model should push the absolute frontier of offensive cyber capability, while the ceiling for open models should rise gradually, because defenders auditing legacy systems need that capability too.

Prices fell; bills did not move much. Input drops from $3 to $2 per million tokens, output from $15 to $10, cache hits from $0.30 to $0.20, a 33 percent cut. But on Artificial Analysis's accounting, one intelligence-index task costs $1.98 versus $2.00 for K3, essentially the same, because output tokens per task rose from 48K to 55K, 14 percent more. So a third cheaper is a statement about unit prices, not about your monthly invoice. Anyone budgeting off the rate card alone will get it wrong.

Three boundaries. First, apart from the Artificial Analysis index, the key figures for finance, hallucination and xdev-bench-hard all come from vendor-run tests and need external replication before they count as settled. Second, safety alignment is not free: on ExploitBench K3.1 Pro scores 74.7, against 91.0 for Claude Opus 5.5 and 100 for GPT-6 Astra, so stripping cyber data buys safety with a real capability concession. Third, the weights only land on November 4, so calling it the strongest open model today is a forward contract. This article does not forecast the landscape after November 4; it records what has been published so far.