Moonshot’s Kimi K3.1 Pro Lands 3 Points From the AI Frontier — With Open Weights Promised by Nov. 4
Key takeaway: Moonshot AI’s Kimi K3.1 Pro, released October 23, 2026, ranks fourth on Artificial Analysis’s Intelligence Index at 55, just three points behind the leading proprietary model. Its full weights are promised by November 4, which would make it the highest-scoring open-weight model yet — but until they appear, it is a model you can use, not one you can download.
Chinese AI lab Moonshot AI has released Kimi K3.1 Pro, the first model in its new K3.1 series and its most capable model in the 3-trillion-parameter class to date. It is available now through Kimi.ai, the Kimi mobile apps, Kimi Work, Kimi Code, and the Kimi API, alongside a faster Highspeed variant that Moonshot says runs two to three times faster than the standard version.
K3.1 Pro is a heavily post-trained upgrade of Kimi K3, the 2.8-trillion-parameter system Moonshot fully open-weighted in July 2026. Artificial Analysis lists K3.1 Pro at the same scale — 2.8 trillion total parameters with 104 billion active parameters in a mixture-of-experts design — with a one-million-token context window and text and image input. The jump here comes from post-training, not from a bigger model.
How close to the frontier is Kimi K3.1 Pro?
On Artificial Analysis’s Intelligence Index, K3.1 Pro scores 55, placing it fourth overall. It trails only Anthropic’s Claude Opus 5.5 at 58, Claude Sonnet 5.5 at 56, and Muse Wonder Full at 56, and sits ahead of OpenAI’s GPT-6 Astra and Google’s Gemini 4 Argon, which each score 53.
The clearest strength is agentic work — multi-step jobs done with tools, rather than single answers. Artificial Analysis measured K3.1 Pro at 67% on Terminal-Bench 4.0, a benchmark for coding agents on realistic terminal tasks, up from just 13% for Kimi K3 and the highest score the evaluator has measured on that test. On GDPval-AA, a simulated knowledge-work evaluation, it scored 1,846 Elo, behind only Claude Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. Its AutomationBench-AA score for agentic software workflows rose from 58% to 73%.
Gains elsewhere were more modest: Humanity’s Last Exam rose from 47% to 52%, CritPt from 23% to 25%, and SciCode from 59% to 61%. On the AA-Omniscience Index, which rewards correct answers and penalises wrong ones, the hallucination rate fell from 53% for Kimi K3 to 44%, with accuracy rising from 48% to 51%.
Vendor numbers vs independent numbers: read both
Moonshot’s own figures tell a similar story with slightly different numbers, and the difference matters. Moonshot reports 65.7% on Terminal-Bench 4.0, against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra. Artificial Analysis, using its own harness, measured 67% for K3.1 Pro. Neither figure is wrong; they come from different evaluation setups, so compare each set within itself rather than mixing the two.
The same caution applies to hallucinations. Moonshot’s internal single-claim test put K3.1 Pro at 8.3%, down from 23.4% for Kimi K3 — a large improvement on that specific stress test, not an overall error rate for everyday use. The independent 44% figure is the more sober guide to how often the model still gets knowledge questions wrong.
| Benchmark | Kimi K3.1 Pro | Comparison |
|---|---|---|
| Intelligence Index | 55 (4th) | Opus 5.5: 58; Sonnet 5.5: 56; GPT-6 Astra: 53 |
| Terminal-Bench 4.0 (independent) | 67% | Kimi K3: 13%; highest measured by Artificial Analysis |
| Terminal-Bench 4.0 (Moonshot) | 65.7% | Opus 5.5: 66.4%; GPT-6 Astra: 58.2% |
| GDPval-AA | 1,846 Elo | Opus 5.5: 1,866; Muse Wonder Full: 1,857 |
| AutomationBench-AA | 73% | Up from 58% for Kimi K3 |
What it is built for, and what it costs
Moonshot pitches K3.1 Pro across coding, 3D creation, financial research, and multi-step agentic tasks — from quick fixes in real codebases to full-project builds and complete financial workflows, plus 3D work across tools such as Unity, Unreal Engine, and Godot. The one-million-token context window supports that pitch, letting users feed in very large codebases or document sets at once, though a large window is no guarantee the model uses all of it well.
Pricing is $2 per million input tokens and $10 per million output tokens, with cache hits at $0.20 per million, according to Artificial Analysis — about half the reported rate card of Claude Opus 5.5. For teams, that combination of near-frontier agentic scores at roughly half the leader’s headline price is the practical story. If the weights arrive as promised, the bigger effect could be that researchers and smaller labs can run and fine-tune a near-frontier model themselves, rather than only renting API access.
Knowns, unknowns, and what to watch
Known: K3.1 Pro is live in Moonshot’s apps and API, scores 55 on the Artificial Analysis Intelligence Index, and leads that evaluator’s Terminal-Bench 4.0 results. Its pricing and one-million-token context window are published.
Unknown: Whether the full weights will actually ship by November 4, and under what licence. Launch analyses noted that no K3.1 licence had been published yet, so the openness of the release cannot be judged until the weights and licence both appear.
What to watch: The November 4 weights release, independent tests of the Highspeed variant, and whether rival labs answer with price cuts or their own open-weight releases. The wider context is a fast-moving Chinese open-model push: just a day earlier, DeepSeek and MiniMax won Beijing’s approval for overseas AI use, and the industry is also converging on shared agent standards, as seen in Anthropic’s donation of Agent Skills to the Linux Foundation.
What this means for you
If you build with AI agents, K3.1 Pro is worth testing now through the API or Kimi Code, especially for terminal-based coding work where its independent scores are strongest. Verify it on your own tasks, and keep the hallucination figures in mind for anything factual. If you are waiting to self-host, do not plan around K3.1 Pro until the weights and licence are actually published — a promised open release is not yet an open release.
Analysis: The headline is less “China catches up” than “the open-weight gap is closing fast.” Artificial Analysis says that gap will shrink from 12 points to 3 once K3.1 Pro’s weights ship. If that holds, the next frontier race may be fought as much on openness and price as on raw scores.
If you found this useful, consider adding USABite as a preferred source in Google Search so our coverage shows up first for you.
Frequently asked questions
What is Kimi K3.1 Pro?
Kimi K3.1 Pro is Moonshot AI’s most capable model in the 3-trillion-parameter class, released on October 23, 2026. It is a heavily post-trained upgrade of Kimi K3, designed for coding, 3D creation, financial research, and multi-step agentic tasks, with a one-million-token context window.
How good is Kimi K3.1 Pro compared to Claude and GPT?
Artificial Analysis ranks Kimi K3.1 Pro fourth on its Intelligence Index at 55, behind Claude Opus 5.5 at 58, Claude Sonnet 5.5 at 56, and Muse Wonder Full at 56, and ahead of GPT-6 Astra at 53. On Terminal-Bench 4.0, it scored 67% in independent testing, the highest score Artificial Analysis has measured on that benchmark.
Is Kimi K3.1 Pro open source?
Not yet. Moonshot AI has promised to publish the full model weights by November 4, 2026, but the weights and licence had not been published at launch. Until they appear, K3.1 Pro is available through Moonshot’s apps and API, but it is not a downloadable open-weight model.
How much does Kimi K3.1 Pro cost?
Kimi K3.1 Pro costs $2 per million input tokens and $10 per million output tokens, with cache hits at $0.20 per million tokens. That is about half the reported rate card of Claude Opus 5.5, and a faster Highspeed variant is also available for more responsive interactions.
Where can I try Kimi K3.1 Pro?
You can try Kimi K3.1 Pro on Kimi.ai, in the Kimi mobile apps, and through Kimi Work, Kimi Code, and the Kimi API. The Highspeed variant is available through the same platforms for users who need faster responses on interactive work.
Sources
- Artificial Analysis: Kimi K3.1 Pro launches at 4th on the Intelligence Index
- CGTN: Kimi K3.1 Pro, an open-weight model within 3 points of the frontier
- Tribune India (ANI): China’s Moonshot AI launches Kimi K3.1 Pro; plans full model weights release by Nov
