A 55 on the independent index, three points off the top
Moonshot AI’s Kimi K3.1 Pro scores 55 on the Artificial Analysis Intelligence Index, fourth overall and three points behind the leader, Claude Opus 5.5 at 58. Moonshot released the model on 23 October 2026: Artificial Analysis dated its launch benchmarks that day, and the-decoder reported the model released and available in Kimi.ai, Kimi Work, Kimi Code and the API. Moonshot’s own launch post carries no publication date, so the release date rests on those two sources.
The index combines several evaluations, and Artificial Analysis runs each model at its maximum effort setting: Kimi K3.1 Pro at “max” effort, and the two Claude entries at “max with fallback”. Opus 5.5 leads on 58, with Sonnet 5.5 and Muse Wonder Full on 56. K3.1 Pro’s 55 sits 11 points above Kimi K3’s 44. On the separate GDPval-AA evaluation it records 1,846 Elo, which Artificial Analysis ranks third.
That three-point framing is index arithmetic at those maximum-effort settings.
2.8 trillion total parameters, 104 billion active
Moonshot describes K3.1 Pro as its “most capable 3T-class open-weight model to date”, with a 1-million-token context window. “3T-class” is the vendor’s rounded label. Artificial Analysis gives the working figure: 2.8 trillion total parameters, 104 billion active. The distance between those two numbers is the signature of a mixture-of-experts model, which routes each token through a fraction of the network. The-decoder cites the same 2.8T and 104B figures for K3.1 Pro, and Moonshot’s Hugging Face page lists Kimi-K3, the previous model, at 2.8T.
Active parameters govern how much of the model runs for each token; total parameters govern the size of the download when the weights appear.
What runs today and what is promised by 4 Nov
K3.1 Pro is live now in Kimi.ai, Kimi Work, Kimi Code and the Kimi API, with a Highspeed variant that Moonshot says runs two to three times faster. The weights are a separate step. Moonshot’s post states that “full model weights will be released by Nov 4, 2026”, and on 26 October the company’s Hugging Face organisation page still listed Kimi-K3 as its most recent repository. No licence has been published, so what reuse the weights would permit is unknown; until terms appear, open weights is the accurate description.
The gap between the best open-weight model and the best proprietary one narrows from 12 points to 3 once the weights ship, according to Artificial Analysis. That narrowing is arithmetic on its index, and it depends on the release happening on time. If the weights slip past 4 November, or arrive under a restrictive licence, the “leading open weights” framing loses its basis.
Pricing is $2 per million input tokens, $10 per million output tokens and $0.20 per million cached input tokens, 33% below Kimi K3 on all three rates. Artificial Analysis and the-decoder publish those figures; Moonshot’s launch post does not contain them and links to a pricing page.
Press reports on 26 October said Moonshot apologised for K3.1 Pro rate limits, reset about 25% of members’ monthly quota and trimmed sub-agents. Moonshot has not confirmed those details. That quota reset is unrelated to the training-data decision below.
Two Terminal-Bench 4.0 numbers that do not reconcile
Terminal-Bench 4.0 is where the vendor’s figure and the independent one part company. Moonshot reports 65.7%, evaluated with the Claude-Code harness, placing it behind Claude Opus 5.5 at 66.4% and ahead of GPT-6 Astra at 58.2%. Artificial Analysis reports 67% and calls that the highest score of any model, ahead of Claude Sonnet 5.5 at 64%. The two published figures differ by 1.3 points, the runs use different comparison sets, and Artificial Analysis does not state a harness for its number. Averaging them would produce a figure neither source supports, and neither has been independently replicated.
Moonshot removed cyber-related training data at two stages
The restriction is Moonshot’s own decision, and the post describes it concretely. Cyber-related content sat inside the general mid-training corpus, and cyber tasks were part of the reinforcement-learning environment mix. Mid-training sits between the initial pretraining run and later fine-tuning; reinforcement learning is the stage that shapes behaviour against task environments. For the release it plans to publish as open weights, Moonshot says it “identified and removed cyber-related data from both training stages, while retaining the general-purpose coding tasks needed for model usefulness”.
The effect is measured on xdev-bench-hard, Moonshot’s in-house benchmark of expert-curated source-code vulnerability discovery and exploitation tasks, drawn from recently disclosed vulnerabilities in software including the Linux kernel, Node.js, JavaScriptCore, Exim and VirGL. A checkpoint without the removals solves 90.5% of problems at pass@3; the final K3.1 Pro reaches 81.0%. Pass@3 credits the model when any of three attempts succeeds, and those runs used the KimiCode harness with no hard timeouts or output token limits.
Those figures compare two checkpoints on one internal benchmark, and they are not a 9.5-point measure of real-world risk. Moonshot has published no task set, specification or leaderboard for xdev-bench-hard, so nobody outside the lab can rerun it as described, and no independent replication has appeared. What exists is a vendor’s own measurement of a reduction it chose to make.
Moonshot says it intends to raise the cyber capability of open weights
Two principles frame the decision, in the company’s words. “First, we aim not to release open weights that would extend the absolute frontier of cyber capability. K3.1 Pro’s offensive cyber capability is, by design, below what our pipeline can produce, and remains behind that of frontier proprietary models.” The post adds that this capability remains behind frontier proprietary models that have been publicly available for at least three months. “Second, we intend to progressively raise the cyber capability available in open weights.”
The second principle is stated intent, with no target, threshold or date, and it sits alongside Moonshot’s line that it will continue to “apply appropriate controls at every training stage”.
The company is describing a deliberate choice to ship less cyber capability than its pipeline can produce, with a stated direction of travel upward. The reduction is measurable only on Moonshot’s own instrument, and the upward path is not measurable from outside. What outsiders can check is the release itself: whether the weights appear by 4 November, and whether anyone beyond the lab can test the same tasks.
