🔍 Read the full analysis: Kimi K3.1 Pro: Near-Frontier Results, Unreleased Weights on ThorstenMeyerAI.com
Get hardware and tech essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Kimi K3.1 Pro became available through Moonshot AI’s apps and API on October 23, with Artificial Analysis scoring it 55.1 and ranking it fourth overall. Its weights are promised by November 4, but are not yet released and the licence has not been published; many of Moonshot’s performance claims come from company tests.
Moonshot AI launched Kimi K3.1 Pro on October 23, and an evaluation by Artificial Analysis placed it fourth on its Intelligence Index, with a score of 55.1. The model is available through Moonshot’s apps and API, but its weights have not been released: the company says they will arrive by November 4, and no licence has been published.
Artificial Analysis reported scores of 57.6 for Claude Opus 5.5, 56.0 for Claude Sonnet 5.5 and 55.8 for Muse Wonder Full, all ahead of Kimi K3.1 Pro. The index score places Kimi close to those systems in that evaluation, but it does not establish equal performance across every task or use case. Artificial Analysis described Kimi as likely to lead open-weight models once the weights are published; that qualification remains material because the files are not yet available.
The evaluator listed K3.1 Pro as a mixture-of-experts model with 2.8 trillion total parameters and 104 billion active parameters, a one-million-token context window, and text and image input. Artificial Analysis recorded its cost at $2 per million input tokens and $10 per million output tokens, or about $2.11 per Intelligence Index task. Its estimate for Claude Opus 5.5 at maximum reasoning was $5.98 per task, though costs depend on the evaluator’s test and usage assumptions.
Artificial Analysis also measured a sharp rise on Terminal-Bench 4.0, a test of agentic coding tasks: Kimi scored 67%, compared with 13% for Kimi K3. On GDPval-AA, a simulated knowledge-work evaluation, it scored 1,846 Elo, behind Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. On Humanity’s Last Exam, Kimi scored 52%, while Opus 5.5 scored 61%.
Infographic 01 · Claimed vs. measured
Moonshot says “matches the strongest.” Independent tests say: close.
The KPIs at the top are independent measurements. Vendor numbers and independent numbers come from different harnesses, so they are shown side by side, never mixed.
Moonshot reports
vendor claims own harness, partly internal tests
| Benchmark | K3.1 Pro | Comparison |
|---|---|---|
| Terminal-Bench 4.0 | 65.7% | vs. Claude Opus 5.5: 66.4% |
| BenchCAD with tools | 92.7% | vs. Claude Opus 5.5: 96.2% |
| OfficeQA Pro v2 | 64.4 | vs. Claude Opus 5.5: 62.2 |
| Kimi Hallu Bench Hard internal | 8.3% | vs. Kimi K3: 23.4% (about −65%) |
Artificial Analysis measured
independent highest reasoning setting per model
| Measurement | K3.1 Pro | Comparison |
|---|---|---|
| Terminal-Bench 4.0 | 67% | vs. Claude Opus 5.5 (max): 60% |
| GDPval-AA | 1,846 Elo | vs. Claude Opus 5.5: 1,866 |
| Humanity’s Last Exam | 52% | vs. Claude Opus 5.5: 61% |
| AA-Omniscience hallucination rate | 44% | vs. Kimi K3: 53% · Gemini 4 Argon: 15% |
The 65% hallucination drop holds on Moonshot’s own stress test. On an independent knowledge test, the rate fell from 53% to 44%.Moonshot: the stress test rates “do not represent overall error rates in daily use.”
Sources: Moonshot AI, Kimi K3.1 Pro tech blog (23 Oct 2026) (left card, vendor charts); Artificial Analysis launch article (23 Oct 2026) and Artificial Analysis model page (data as of 24 Oct 2026) (KPIs and right card).
Powered by Thorsten Meyer AI
A Strong Score, With Access Still Limited
The result matters because high-performing models with downloadable weights can give organisations more control over deployment, data handling and customisation than a service available only through a provider. But Kimi K3.1 Pro is not yet downloadable, so users cannot currently verify how it performs on their own hardware or inspect the released model files. A promised release is not the same as an available open-weight model.
Moonshot’s listed API prices are also relevant to developers comparing hosted options. Artificial Analysis estimated a lower cost per index task than for Opus 5.5 in its evaluation, while the models’ benchmark performance differed. Those figures are not a guarantee of savings in production: actual spending depends on prompt size, generated tokens, reasoning settings and the tasks being run. The reported lead in Terminal-Bench suggests a notable improvement over Kimi K3 for that specific agentic coding test, not a universal advantage.
There is also a distinction between benchmark strength and independent verification. Artificial Analysis supplies an outside comparison on selected tests; many eye-catching examples in Moonshot’s launch materials come from the company’s own demonstrations and internal evaluations. Companies considering the model will need the weights, licence terms and further testing before judging whether it fits their security, legal and technical requirements.
GPU workstation for running large language models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What the Benchmarks Actually Measure
Kimi K3.1 Pro is described as a more heavily post-trained version of Kimi K3. It went live on October 23 across the Kimi apps, Kimi Work, Kimi Code and the Kimi API. The launch therefore provides access to a hosted model now, even though the announced release of weights is still pending.
Moonshot says K3.1 Pro can build projects across Unity, Unreal Engine and Godot, create parametric 3D models in Blender and CAD software, and perform finance-related work. The company reports a 92.7% result on BenchCAD with tools, compared with 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra. These are claims and results presented by Moonshot; the source material does not establish that the demos or internal tests have been independently reproduced.
Moonshot also says its internal Kimi Hallu Bench Hard test found claim-level hallucinations fell from 23.4% for Kimi K3 to 8.3% for K3.1 Pro, a reduction it describes as about 65%. Artificial Analysis measured a smaller change on a different evaluation: its AA-Omniscience hallucination rate fell from 53% to 44%. The measures use different tests and should not be treated as interchangeable. Moonshot itself says its stress test is deliberately designed to provoke errors and does not represent overall error rates in everyday use.
“The test is deliberately crafted”
— Moonshot AI
high VRAM graphics card for AI model inference
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Weights, Licensing and Claims Pending
The weights have not been released, and the licence governing any eventual download or reuse remains unpublished. Until both are available, developers cannot determine from the release whether they can run the model locally, modify it, or use it under their intended commercial terms. Artificial Analysis’s description of Kimi as a future leading open-weight model depends on those details and on the promised release taking place.
It is also unclear whether Moonshot will meet its November 4 target, what hardware and operating costs would be required for self-hosting, and whether independent teams can reproduce the company’s demos. A 2.8-trillion-parameter model may present substantial infrastructure demands; the source material does not provide a verified deployment specification. Benchmark results cover defined tests, not every real-world task, and the supplied information does not resolve how the model performs across different languages, industries or safety conditions.
AI development workstation with 128GB RAM
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The November 4 Release Deadline
The next stated milestone is Moonshot’s planned November 4 release of the weights. If they appear, readers and developers can review the files and licence, assess whether the model is practical to run, and compare independent results with the current hosted benchmarks. The launch information does not specify a release time or provide a licence in advance.
Until then, Kimi K3.1 Pro can be accessed through the listed Moonshot products and API at the stated token prices. Further evaluation should distinguish results from Artificial Analysis’s published tests from Moonshot’s internal benchmarks and demonstrations. Any changes to the promised schedule, licence or independent rankings remain to be seen.
As an affiliate, we earn on qualifying purchases.
Key Questions
Is Kimi K3.1 Pro open-weight now?
No. The model is available through Moonshot’s apps and API, but its weights have not been released. Moonshot says they are planned for November 4.
How did Kimi K3.1 Pro rank in independent testing?
Artificial Analysis gave it a 55.1 Intelligence Index score and ranked it fourth among the models listed. Claude Opus 5.5, Claude Sonnet 5.5 and Muse Wonder Full scored higher in that comparison.
What does Moonshot claim the model can do?
Moonshot says K3.1 Pro can work on game projects, create parametric 3D models and perform finance tasks. Those examples and several related results come from the company’s own demonstrations or internal tests, rather than independent confirmation in the supplied material.
Has Moonshot published a licence for the weights?
No licence was published in the information available for this report. The terms will matter to anyone assessing modification, redistribution or commercial use.
Does the benchmark score show it will outperform other models for every task?
No. The Intelligence Index and other scores measure performance on specified evaluations. They do not establish that Kimi K3.1 Pro will be better for every workload, and real-world costs and results can vary by task and settings.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
