Companies choosing models for coding agents get a near-frontier option at US$2 per million input tokens that they can run on their own servers once the weights arrive November 4.
Moonshot AI, the Beijing developer of the Kimi chatbot, released Kimi K3.1 Pro in the early hours of October 24, Hong Kong time. The model is a post-trained upgrade of K3 with a 1-million-token context window, and it comes with a faster companion, K3.1 Pro Highspeed, which the company says runs two to three times quicker. Moonshot’s launch post says both are live in the Kimi apps, Kimi Work, Kimi Code and the Kimi API, and that the full model weights will be published by November 4, 2026.
The headline claim is in coding agents. On Terminal-Bench 4.0, which tests AI agents carrying out tasks in a computer terminal, Moonshot says K3.1 Pro scored 65.7%, against 66.4% for Anthropic’s Claude Opus 5.5 and 58.2% for OpenAI’s GPT-6 Astra. Every benchmark figure in the launch post is reported by Moonshot itself and none has been replicated independently, as The Next Web also points out. The one outside score so far comes from Artificial Analysis, a benchmarking firm, which ranks the model fourth on its Intelligence Index with 55 points, behind Opus 5.5, Claude Sonnet 5.5 and Meta’s Muse Wonder Full, according to The Decoder and a briefing from The Information.
The upgrade is also cheaper. The Decoder reports API prices of US$2 per million input tokens and US$10 per million output tokens, down from US$3 and US$15 for K3, a cut of about a third; Moonshot’s OpenRouter listing shows the same rates, with the Highspeed version at US$4 and US$20.
Moonshot spent its extra training on agents, finance and 3D design
K3.1 Pro keeps K3’s design: 2.8 trillion parameters in total, of which 104 billion are used for each piece of text, The Decoder reports. Moonshot calls it "our most capable 3T-class open-weight model to date" and says the gains come from substantially more post-training compute, the stage where a base model is tuned for specific tasks. On Terminal-Bench 4.0, K3 scored 13%, according to The Decoder. On Moonshot’s internal long-horizon coding test, the score rose to 61.2% from 55.95% for K3.
The launch post puts weight on work outside coding. On BenchCAD, a test of building 3D models with design software, K3.1 Pro scored 92.7 with tools, below Opus 5.5 at 96.2, and 72.8 without tools, against 73.0 for Opus 5.5. Moonshot says its 64.4 on OfficeQA Pro v2 is the best of any model it tested, and its internal Financial Research score rose to 70.4 from 58.8. The Decoder adds that Moonshot launched a 3D Make feature on Kimi.com for previewing and exporting models.
Moonshot also reports fewer invented facts. On the hard version of its own Kimi Hallu Bench, the share of false claims fell to 8.3% from 23.4% for K3, while GPT-6 Astra scored 5.5% on the same test. The company describes the test as a deliberately adversarial stress test, so the rates do not describe everyday error levels.
Moonshot cut the model’s hacking skills before the weights go public
Moonshot says it removed cyber-related training data from the later training stages of the open-weight release while keeping general coding data. On its internal xdev-bench-hard, which tests exploiting real software flaws, a version trained without those cuts solved 90.5% of tasks; the released model solved 81.0%, against 85.7% for DeepSeek-V4.1-Flash and 33.3% for K3. The company writes that "we aim not to release open weights that would extend the absolute frontier of cyber capability," and says K3.1 Pro trails proprietary models that have been public for at least three months.
The choice matters because downloaded weights cannot be recalled, The Next Web notes. Moonshot also faces scrutiny in Washington: in July the White House accused it of using banned Nvidia chips and copying an Anthropic model to build K3, without publishing evidence, and in August K3 escaped a test sandbox to look up a benchmark’s answers online, according to The Next Web.
The November 4 weights will let outsiders test Moonshot’s claims
The launch had been trailed. On October 23, The Paper, citing Kechuangban Daily, reported that Kimi’s official accounts had posted a short video labeled "Model: Kimi K3.1." Business and Enterprise customers must have an administrator switch the new model on, because it is off by default at launch, the launch post says.
If the weights arrive on schedule, K3.1 Pro would become the top open-weight model on the Artificial Analysis index and narrow the gap to the best proprietary model from 12 points to 3, The Decoder calculates. Moonshot is raising money as it ships. Yicai reported in July that it had raised US$3.5 billion at a US$35 billion valuation and was seeking a new round at about US$50 billion. In September, TechNode reported that Moonshot had submitted a confidential listing application in Hong Kong; the company had no comment. Domestic rival DeepSeek is also raising, as it nears a RMB100 billion round.
The first concrete test is the weights themselves: once they appear, outside labs can run Terminal-Bench and the cyber tests on their own hardware, and their numbers will show how far Moonshot’s figures hold up.