Kimi K3.1 Pro: 55 on the AI index, 3 points behind Opus 5.5
Moonshot AI says its new Kimi K3.1 Pro is the best open-weights model anyone has built. The weights are not out yet — the company says they land by November 4, and if that date holds, the gap between the best open model and the best proprietary one shrinks from twelve index points to three.
Kimi K3.1 Pro scores 55 on the Artificial Analysis Intelligence Index, fourth overall and three points behind Claude Opus 5.5's 58, according to the evaluator's launch analysis. The model is a post-trained version of Kimi K3 with what Moonshot calls significantly more post-training compute, and it keeps the same skeleton: 2.8 trillion total parameters, 104 billion active, a one-million-token context window, and text plus image input. The jump is eleven points over K3 on the same index, and enough to put a model anyone can eventually download within rounding distance of Anthropic's flagship.
That framing is the story, not the score. For two years the working assumption has been that open weights trail the frontier by a generation and catch up six months later at a discount. K3.1 Pro compresses the lag to a single model version. If the weights ship on schedule it becomes the highest-scoring open model available by a wide margin, and the price of running a near-frontier model stops being a function of what a US lab decides to charge.
Where the gains actually are. Almost all of the improvement sits in agentic work, which is where the money is. On GDPval-AA, the index of economically valuable tasks, K3.1 Pro reaches 1,846 Elo — a 312-point climb over K3 — good for third behind Claude Opus 5.5 at 1,866 and Meta's Muse Wonder Full at 1,857. On AutomationBench, which tests agentic SaaS workflows, it goes from 58% to 73%. On Terminal-Bench 4.0, the terminal-agent test that most labs now treat as the practical proxy for "can it do my job," Kimi K3 went from 13% to 65.7% in one post-training pass.
That last number is the one to sit with. A model that could barely drive a terminal now sits within a point of Opus 5.5. Artificial Analysis measured it even higher, at 67%, which it reports as the best score of any model it has tested and ahead of Claude Sonnet 5.5's 64%. The two figures differ because harness settings differ; the direction does not.
The rest of the eval sheet is more pedestrian, and the honesty is worth noting. Humanity's Last Exam rises from 47% to 52%, CritPt from 23% to 25%, SciCode from 59% to 61%. Those are incremental gains on hard reasoning, not breakthroughs. The hallucination numbers moved more: on Artificial Analysis's Omniscience index the score climbs from 20 to 30 as the hallucination rate falls from 53% to 44% and accuracy rises from 48% to 51%. Moonshot reports a steeper drop on its own internal stress test, from 23.4% to 8.3% claim-level hallucination, but says plainly that this is a stress test and not a measure of general error rates.
One caution on the headline comparison. Moonshot's eleven-point gain is measured against K3's current 44 on that index, while K3 was widely reported at 57 when it launched in July. Some part of the apparent jump is the index being revised and re-based since, not new capability — worth remembering before anyone quotes the eleven points as pure progress.
The price cut is the part enterprises will feel. Token prices fall 33% across the board: $2 per million input tokens, $10 per million output, $0.20 for cache hits, all three down exactly a third from K3's $3/$15/$0.30, per Moonshot's launch post. A Highspeed variant runs two to three times faster at double the price, $4/$20. Both are live now through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Cost per task lands at $1.98 on the Intelligence Index, essentially flat against K3's $2.00, because the cheaper tokens are partly eaten by the model generating 14% more output — 55,000 output tokens per task against 48,000. That combination is what puts it on the Pareto frontier: no model at a comparable parameter count buys more capability per dollar. For a team running agents in a loop, where token spend compounds across thousands of calls, a third off output pricing at three points below the frontier is the whole argument.
The open-weights trade-off nobody will lead with. Moonshot devotes an unusual chunk of its launch post to cybersecurity, and the disclosure is more revealing than the benchmark table. The company says vulnerability-research capability kept climbing as it scaled post-training, because cyber content sat in the general mid-training corpus and related tasks were in the RL mix. For the open release it stripped cyber data from both phases.
The cost of that scrub is measurable. On Moonshot's internal xdev-bench-hard, which reproduces recently disclosed bugs in the Linux kernel, Node.js and JavaScriptCore, the uncleaned model solves 90.5% of tasks. The cleaned version solves 81.0%. Kimi K3 solved 33.3%. DeepSeek-V4.1-Flash, which made no such commitment, reaches 85.7% on the same test.
Read those four numbers in order and the trade-off is obvious: the safest open release in the family is now the weakest open model on exploit work, behind a competitor that did not scrub. Moonshot's stated position is that no open-weights model should push the absolute frontier of offensive cyber, while open models still need to be useful for auditing legacy systems. That is a defensible line — Anthropic has already shown that even GLM-5.3, a mid-table model, can write exploits near its own frontier models. But it also means the open-weights race now has a capability ceiling that vendors set unilaterally, and buyers cannot audit where it was drawn.
The counter-argument. Three points on a composite index is inside the noise band, and the sub-scores say so: K3.1 Pro wins SWE Marathon, BrowseComp and OfficeQA Pro v2, loses BenchCAD with tools (92.7% to Opus's 96.2%) and loses badly on security work, where ExploitBench puts it at 74.7% against Opus 5.5's 91.0% and GPT-6 Astra's 100%. Moonshot's own summary concedes that GPT-6 Astra still hallucinates less. And the deployment problem is unchanged: 2.8 trillion parameters means almost nobody will self-host this. The weights matter because they let clouds, sovereign programs and well-funded labs run it on their own terms, not because a startup will put it on a laptop.
We covered the tier-level preview in September — Kimi K2.8 Preview puts near-K3 coding in every membership tier — and the pattern has held since: Moonshot ships the consumer product first, then the API, then the weights, each step cheaper.
What to watch. November 4 is the only date that matters. If the weights land, the open-versus-proprietary gap on this index is three points and the pricing argument for closed frontier models gets harder to make. If they slip past November — K3's weights were promised and delivered eleven days after launch — the fourth-place score is just another number on a leaderboard nobody can download. Watch the license terms too: Kimi's K3 license carried commercial conditions, and a permissive K3.1 license would be a bigger competitive event than the index score.
Is a three-point gap small enough that you would move a production workload off a proprietary API to an open model you host yourself? Write to us via our contact page.
