Kimi K3.1 Pro: an open-weight model within 3 points of the frontier
Moonshot AI released Kimi K3.1 Pro on Friday, and it debuted fourth on the Artificial Analysis Intelligence Index with a score of 55, trailing only Anthropic's Claude Opus 5.5 and Sonnet 5.5 and Muse Wonder Full.
Once the Chinese startup publishes the full weights, expected by November 4, K3.1 Pro will become the highest-scoring open-weight model ever released, narrowing the gap between the leading open-weight and proprietary systems from 12 points to 3, according to Artificial Analysis.
The gap between leading open-weight and proprietary systems narrows from 12 points to 3 — but the gap that governs purchasing is measured elsewhere.
The model is a post-trained version of Kimi K3, the 2.8-trillion-parameter system Moonshot fully open-weighted in July, which was then the largest open-weight model in existence. K3.1 Pro retains the 3-trillion-parameter class and a 1-million-token context window.
The new model's biggest gain is on Terminal-Bench 4.0, which tests coding agents on real terminal tasks: K3.1 Pro scores 65.7%, ahead of OpenAI's GPT-6 Astra at 58.2% and close to Claude Opus 5.5 at 66.4%. Pricing is $2 per million input tokens and $10 per million output, half the rate card of Claude Opus 5.5.
The release drew roughly 850 points on Hacker News, a score that reflects both upvotes and comment activity, with nearly 500 comments by Friday evening. The most discussed trait was temperament. "Kimi is so much more blunt and argumentative than USA models," one highly visible comment read. "It's so funny, it will say things like, 'We've already been over this 3 times and you refuse to accept my advice, asking the same question another way won't change the answer.'"
The unresolved question is whether K3.1 Pro's benchmark proximity to Claude Opus 5.5 translates into real-world equivalence. OrcaRouter notes two caveats: the Terminal-Bench scores were produced at different reasoning settings, and the two-time price advantage is per token, not per finished task. Artificial Analysis puts K3.1 Pro's cost per Intelligence Index task at $1.98, roughly equal to K3's $2.00, because it generates 55 thousand output tokens per task – 14% more than K3.
China is rising in open-weight AI. Hugging Face's 2026 report found that open models developed in China accounted for 41% of the platform's downloads, ahead of the US at roughly 35-36.5%. Science and Technology Daily wrote that Kimi K3 "has sent shockwaves through the global tech community," and Axios bluntly noted that "America's commanding lead in advanced AI is gone."
Moonshot, founded in 2023, went from a long-context chatbot to a $50 billion valuation in under four years.
Article: 65.7% on Terminal-Bench 4.0. That is Moonshot's own number for Kimi K3.1 Pro, and it sits less than a point behind Claude Opus 5.5 at 66.4% — from a model whose full weights are promised by November 4.
Artificial Analysis, running its own harness, scored it higher still: 67.2%, ahead of Claude Sonnet 5.5. Fourth on the Intelligence Index at 55, behind only Opus 5.5 at 58 and the pair of Sonnet 5.5 and Muse Wonder Full at 56. Per-token pricing is $2 per million input and $10 per million output, half of Opus 5.5.
This is the first time a self-hostable model has been within a month of the closed frontier on agentic coding. It is also not the coronation those numbers imply.
Last week's issue covered DeepSeek's agent-cost floor and Meta's return to US open weights — the price shock. This one is a different benchmark: near-frontier quality, not near-zero cost. The delta matters because it changes the question from "can open weights be cheap?" to "will anyone actually deploy them?"
Start with what actually improved, because the gains are not uniform. Terminal-Bench 4.0 went from 12.6% on Kimi K3 to 67.2% — a 54.6-point jump. GDPval-AA rose 312 Elo, AA-Briefcase 179, AutomationBench 15 points. Kimi's own long-horizon test had the model build a human walking simulator, diagnose 36 errors in shoulder-muscle moment-arm calculations, and get the simulated body walking for 11.2 seconds.
Then look at knowledge. SciCode gained 1.3 points. CritPt gained 1.2. Humanity's Last Exam, 5.6. The long-context eval actually fell 2.4 points. Artificial Analysis read this as a model taught to work rather than a model taught more: RL across hundreds of domains sharpens multi-step execution and leaves raw recall where it was.
That is a coherent story, and it is the story of a specialist. Enterprise procurement does not buy specialists by accident.
Then the cost. Per-token, K3.1 Pro is a third cheaper than K3 — a genuine cut. Per task, it is $1.98 against K3's $2.00. Roughly flat. The reason is that K3.1 Pro emits 55,000 output tokens per Intelligence Index task, 14% more than K3's 48,000. The rate card halved; the bill did not.
That gap between advertised price and realized cost is exactly the gap that benchmarks don't measure, and it is where the substitution case starts to wobble. The sibling piece on verbosity-as-hidden-bill covers the mechanics; here the point is narrower. A buyer comparing $2/$10 to Opus 5.5's rate card is comparing a price list, not a deployment.
The launch reception reflected that. Roughly 850 points on Hacker News, nearly 500 comments by Friday evening — and the most-discussed trait was not coding. It was temperament. "Kimi is so much more blunt and argumentative than USA models," one highly visible comment read. Another quoted the model back: "We've already been over this 3 times and you refuse to accept my advice, asking the same question another way won't change the answer."
That is a real product property, and it is not on the Artificial Analysis table. It also isn't disqualifying. But it tells you the community's first instinct was to evaluate the model as an interlocutor, not as a component — and those intuitions are what propagate into procurement conversations months later.
The release was also shadowed by allegations of undisclosed paid promotion. One person reportedly tweeted a screenshot of press instructions telling recipients not to disclose that the posts were paid; per Interconnects, no one corroborated it afterward. Nathan Lambert's read: bad form, a mess, no company should do it. The promotional-misconduct question is separate from capability — but it lands in the same trust ledger buyers draw on.
And there is the cyber profile, deliberately weakened. Moonshot removed cyber-related training data from the post-training set; on its internal exploit benchmark the unfiltered checkpoint solved 90.5% of tasks against 81.0% for the shipped model. Then the shipped model scored 64% on Artificial Analysis' Cyber Index anyway — the highest publicly accessible result, second overall behind trusted-access GPT-6 Sol Daybreak Blue at 69%.
Sit with that for a second. Moonshot tried to make the model worse at security work and it still topped the public leaderboard. That is a capability fact, and it is also a governance fact: the version enterprises can actually self-host is the safety-filtered one, which is not the version the benchmarks ran on. OrcaRouter's caveat cuts the same direction: the Terminal-Bench figures came from different reasoning settings.
Now the procurement reality, which is the part that decides this. Buyers are not benchmarking. They are deploying, and deployment consumes a different set of artifacts.
A license. Moonshot has not published redistribution terms, commercial-use restrictions, checkpoint precision, or full inference requirements. "Soon-to-be open-weight" is the accurate phrase; "open source" is not. The 10/19 issue asked whether enterprises would accept Meta's attribution and revenue-triggered license as procurement friction. K3.1 Pro's still-unpublished license is the next test of that same question.
A serving stack. 104B active parameters is not a 104B dense model. Deployment has to hold the expert pool plus runtime state: roughly 5.6 TB at 16-bit, 1.4 TB at 4-bit, before quantization metadata and serving overhead. Self-hostable and consumer-device-local are different properties, and nobody calling this a local model is doing the buyer any favors.
A harness. Kimi is shipping Kimi Hosted Agents — a managed harness with a 24/7 execution environment for long-running and multi-agent work. That is the right product to ship, and it is also the tell. The model alone doesn't get you the deployment; the harness does, and Moonshot's harness is Moonshot's.
A security posture. A filtered checkpoint, a public Cyber Index crown, and an unresolved moderation story are three facts that a CISO has to reconcile before signing.
A vendor. Moonshot went from a long-context chatbot to a $50 billion valuation in under four years. That is an extraordinary run and it is not the same thing as a decade of enterprise support relationships.
None of this means K3.1 Pro fails commercially. It means the substitution calculus did not move as far as the benchmark gap did. The gap narrowed from 12 points to 3 — but the gap that governs purchasing is measured in licenses, serving footprints, harness ownership, and vendor trust, and on those axes the distance is roughly what it was.
The open-weight story this month is real. The closed labs should read it as pressure on price and on release cadence. The open-weight camp should not read it as a procurement victory, because the thing that got faster was the model. Everything downstream of the model is still the hard part.