Moonshot AI launched Kimi K3.1 Pro on October 23, 2026, a post-trained upgrade of its K3 model that scores 65.7% on Terminal-Bench 4.0, just 0.7 points behind Anthropic's Claude Opus 5.5 at 66.4% and 7.5 points ahead of OpenAI's GPT-6 Astra at 58.2%. The model ships today through Moonshot's API at $2 per million input tokens; the full open weights follow on November 4.
The release matters for two reasons that go beyond the benchmark table. First, it is the strongest open-weight model anyone has shipped to date, and the gap to the closed frontier has shrunk to less than a percentage point on agentic coding. Second, Moonshot deliberately stripped cyberattack training data out of the model before the open release, a safety precedent other open-weight labs will now be measured against.
The benchmark ledger
Every number below comes from Moonshot's launch materials, run at maximum reasoning effort, mostly inside the Claude Code harness. They are vendor-reported, not independently verified, and rival scores mix Moonshot's own runs with official leaderboards. Read them as Moonshot's case, not a neutral audit.
The Terminal-Bench 4.0 result is the headline: 65.7% puts K3.1 Pro ahead of GPT-6 Astra (58.2%), Fable 5.1 (57.9%), DeepSeek V4.1 Flash (42.3%), and GLM-5.3 (41.8%), and 0.7 points behind Claude Opus 5.5 (66.4%). On SWE-Marathon it actually leads at 51.7, ahead of Opus 5.5 at 46.3. But the ledger has an honest gap: on Terminal-Bench Science 0.1, K3.1 Pro scores 38.1 against 68.1 for Astra and 63.3 for Opus 5.5. Moonshot's model is a coding specialist that trails badly on science tasks, and the company published that weakness itself.
| Benchmark | Kimi K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 65.7% | 66.4% | 58.2% |
| SWE-Marathon | 51.7 (leads) | 46.3 | not reported |
| Terminal-Bench Science 0.1 | 38.1 | 63.3 | 68.1 |
| BenchCAD, with tools | 92.7 | 96.2 | 95.9 |
| BenchCAD, without tools | 72.8 | 73.0 | not reported |
| OfficeQA Pro v2 (office work) | 64.4 (SOTA claim) | not reported | not reported |
| Kimi Hallu Bench Hard (false-claim rate) | 8.3% | not reported | 5.5% |
The independent check comes from Artificial Analysis, which scores K3.1 Pro at 55 on its Intelligence Index, fourth overall behind only three closed models (Claude Opus 5.5, Claude Sonnet 5.5, and Muse Wonder Full) and 11 points above K3. That is the number to watch when the open weights land and independent labs can run their own harnesses.
The price cut and the two-week gap
K3.1 Pro costs $2 per million input tokens and $10 per million output tokens on the Kimi API, with cached input at $0.20. That is a 33% cut from K3's pricing. A Highspeed variant runs two to three times faster at $4 and $20. Three reasoning levels (low, medium, max) let buyers trade cost against effort.
The notable business decision is the calendar: API access now, open weights November 4. Moonshot gets roughly two weeks to monetize the hosted version before anyone can self-host. It is the same playbook we mapped in our open-weight benchmark wars analysis: the weights are the moat-breaker, so the lab sells the head start first.
The cyber scrub
Moonshot found cyber-related content in K3's mid-training data and cyber tasks in its reinforcement-learning mix, and removed both for the open-weight release. On its internal exploit benchmark, the unfiltered checkpoint solved 90.5% of tasks at pass@3; the shipped model solves 81.0%. Moonshot's stated rule is that it will not ship open weights that extend the frontier of offensive cyber capability.
This is a genuine precedent. Open-weight releases have historically shipped whatever the lab trained, and safety was the downloader's problem. Moonshot is asserting a duty of care at the weights level, and publishing the before/after numbers to prove the scrub cost real capability. Whether rivals follow will decide if this becomes a norm or a marketing line.
What the announcement does not establish
Moonshot's benchmarks are self-reported at maximum effort with a harness choice (Claude Code) that favors agentic coding. The 0.7-point gap to Opus 5.5 could easily invert under independent evaluation, and the science-benchmark deficit suggests the model is narrower than the headline implies. The November 4 weights date is a promise, not a delivery, and no license terms have been specified yet. Until independent labs run the weights, treat every number here as Moonshot's opening bid.
The open-weight playbook, explained
Why do labs release weights at all, if the API is the business? Because weights buy what APIs cannot: researcher goodwill, ecosystem lock-in, and regulatory cover. A lab that publishes weights can claim it is democratizing AI while still charging for the convenient hosted version. The two-week API-first gap is the tell: it prices the convenience premium before the weights commoditize it. For buyers, the durable question is not who tops this week's benchmark but whose weights you can actually run, under what license, at what hardware cost. K3.1 Pro's 2.8-trillion-parameter size means self-hosting remains a datacenter-scale affair, which is exactly why the API head start is worth pricing.
By ByteHerald Staff. Moonshot AI announced Kimi K3.1 Pro on October 23, 2026, reporting Terminal-Bench 4.0 at 65.7% versus 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra, with open weights promised by November 4 and API pricing of $2/$10 per million tokens. Artificial Analysis independently scored the model at 55 on its Intelligence Index, fourth overall and 11 points above K3, as reported October 23, 2026. Moonshot disclosed removing cyber training data for the open release, with its internal exploit benchmark falling from 90.5% to 81.0% after the scrub.
Comments
Post a Comment