Moonshot AI launched Kimi K3.1 Pro on October 23, 2026 through its applications and API, describing it as a model intended for a subsequent open-weight release. It promised downloadable weights by November 4. At launch, the weights were not yet published, and the reviewed launch materials did not specify their licence. Moonshot calls K3.1 Pro its most capable 3T-class open-weight model to date; until the download exists, that description is a schedule, not a property of the model.
What did ship is a post-training upgrade, not a new base model. K3.1 Pro is Kimi K3 retrained with substantially larger post-training compute, including reinforcement learning run jointly across hundreds of downstream domains. Moonshot also released a Highspeed variant it says runs two to three times faster, and cut the standard model's token prices by a third, to $2 per million input tokens and $10 per million output.
The independent score, and the two Terminal-Bench numbers
The strongest evidence in this launch is not Moonshot's. Artificial Analysis, which runs its own evaluations, scores K3.1 Pro at 55 on its Intelligence Index in its launch-day evaluation of October 23, fourth overall behind Claude Opus 5.5 at 58 and Claude Sonnet 5.5 and Muse Wonder Full at 56, and 11 points above K3's 44. Leaderboards move: the ranking is specific to Artificial Analysis's methodology and to launch day. Its agentic results moved the most: GDPval-AA rose 312 Elo points to 1,846, third overall, and AutomationBench-AA rose from 58% to 73%. Gains elsewhere were smaller, with Humanity's Last Exam up from 47% to 52% and SciCode from 59% to 61%.
Read the coding headline carefully, because there are two of them. In Artificial Analysis's own Terminal-Bench 4.0 run, K3.1 Pro scored 67% — the highest score that evaluator had measured as of its October 23 launch report. In Moonshot's own run, it scored 65.7%, narrowly behind Claude Opus 5.5 at 66.4%, a comparator reported in Moonshot's evaluation materials. Moonshot's footnotes say its model was evaluated in a Claude Code harness and that the remaining scores were taken as the maximum reported in their blog posts or the official Terminal-Bench 4.0 leaderboard — imported, not rerun. Those are different runs under different conditions, and treating either figure as a head-to-head result misreads both.
A price cut that does not lower the bill per task
The 33% price cut is real, but token prices are not task costs. Artificial Analysis reports that K3.1 Pro generated approximately 55,000 output tokens per Intelligence Index task, 14% more than K3. Under that evaluator's settings and workload, the measured cost per task remained roughly unchanged at $1.98 — a figure for Artificial Analysis's Index tasks, not a typical cost for any given user task. Anyone budgeting a migration on the headline price cut should budget on cost per successful task instead.
Moonshot's own accuracy claims need the same discipline, because there are three hallucination numbers in this launch and they come from three different tests. On its internal Kimi Hallu Bench Hard stress test, the company reports the claim-level hallucination rate falling from 23.4% for K3 to 8.3%, and it states plainly that the test is a deliberate stress test whose rates do not represent daily error rates. Its API quickstart describes a different evaluation, built from de-identified conversations in which users had flagged factual errors in earlier models' responses, with a different result again: the claim-level hallucination rate fell from 23.35% for K3 to 6.85%. Artificial Analysis's separate AA-Omniscience measurement has the hallucination rate falling from 53% to 44%. Three sources, three different tests and three different pairs of numbers. Quote any of them only with the test named.
The cyber disclosure is the unusual part
For the open-weight release, Moonshot says it found cyber-related content embedded in its mid-training corpus and cyber tasks in its reinforcement-learning environment mix, and removed both. On its internal xdev-bench-hard exploit benchmark, the checkpoint without those removals solves 90.5% of tasks at pass@3, according to Moonshot; Moonshot reports the final K3.1 Pro at 81.0%. The 9.5-percentage-point difference is a reported difference, not an independently established safety improvement, and every figure in this paragraph is Moonshot's own, from an internal benchmark no outside group has run. Moonshot's stated principle is that it aims not to release open weights that would extend the absolute frontier of cyber capability. The disclosure provides a specific, testable claim about how Moonshot says it changed the model before the planned weight release, and it gives the November release something specific to be checked against. It does not independently establish the effectiveness of those changes or the model's broader cybersecurity risk.
What November 4 has to deliver
Artificial Analysis puts the conditional plainly: once the weights are published, K3.1 Pro becomes the highest-scoring open-weight model on its Intelligence Index, and the gap between the leading open-weight and proprietary models on that index narrows from 12 points to 3. Those are Artificial Analysis's index calculations, not a universal measure of the capability gap. The claim also rests on two undelivered items: the weights themselves, and a licence whose terms nobody has seen. A restrictive licence would not necessarily change the model's benchmark ranking or the fact that its weights are downloadable. It could substantially change who can use those weights, for what purposes, and under what conditions — as the MiniMax Community License does for MiniMax's open weights, where commercial use is conditional and products or services generating more than US$20 million in yearly revenue need prior written authorization from MiniMax. The ranking is launch-day and independently measured. The openness is pending.
Sources
- [1] Moonshot AI, “Introducing Kimi K3.1 Pro” tech blog (official; benchmark footnotes, cyber data-removal disclosure, Hallu Bench Hard figures, weights by Nov 4, 2026) — read first-hand October 23, 2026Read source
- [2] Kimi API Platform, “Kimi K3.1 Pro” quickstart (official; $1 top-up access condition, reasoning-effort settings, de-identified-conversation hallucination evaluation) — read first-hand October 23, 2026Read source
- [3] Artificial Analysis, “Kimi K3.1 Pro launches at 4th on the Intelligence Index” (October 23, 2026; independent Index score of 55, its own Terminal-Bench 4.0 run at 67%, pricing $2/$10/$0.20, cost per task $1.98, weights not published) — read first-hand October 23, 2026Read source
- [4] Kimi (Moonshot AI) official LinkedIn launch post (timestamped 2026-10-23T17:22Z; 2.8T/104B, 1M context, live channels, open weights by Nov 4, 2026)Read source