Kimi K3.1 Pro’s Strong Showing: Three Points Off The Frontier

🔍 Read the full analysis: Kimi K3.1 Pro’s Strong Showing: Three Points Off The Frontier on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Moonshot AI’s Kimi K3.1 Pro launched on Oct. 23 and scored 55.1 on Artificial Analysis’s Intelligence Index, ranking fourth in the figures reported the following day. That puts it 2.5 points behind the highest-scoring model listed, Claude Opus 5.5 at 57.6; although Moonshot describes K3.1 Pro as an open-weight model, its weights have not yet been released and the company has not published a licence for them.

Artificial Analysis lists K3.1 Pro as a mixture-of-experts model with 2.8 trillion total parameters and 104 billion active parameters. It supports text and image input and has a reported one-million-token context window. The model is available through the Kimi apps, Kimi Work, Kimi Code and the Kimi API, according to the source report.

In Artificial Analysis’s comparison, K3.1 Pro scored 55.1, behind Opus 5.5 at 57.6, Claude Sonnet 5.5 at 56.0 and Muse Wonder Full at 55.8. The gap to the listed leader is 2.5 points, often rounded to about three. Artificial Analysis’s assessment is conditional: it says K3.1 Pro would lead open-weight models once the weights are published. The current results do not mean users can already download and run those weights.

The listed API price is $2 per million input tokens and $10 per million output tokens. Artificial Analysis estimates about $2.11 per Intelligence Index task for K3.1 Pro, compared with $5.98 for Opus 5.5 at its maximum reasoning setting. These task-cost figures reflect the evaluator’s testing and are not a universal measure of what every use will cost.

A Narrower Gap, With Caveats

The result matters because it places a model that Moonshot says will have open weights close to several leading closed models on a broad benchmark. If the weights are released under usable terms, organizations could have another option for deploying a high-scoring model on their own infrastructure rather than relying only on a hosted service. That could matter for data handling, customization and dependence on a provider, though the benchmarks alone do not establish how well the model will perform in a particular company’s workflow.

Cost is another point of comparison. Artificial Analysis’s estimate puts K3.1 Pro’s cost per index task below Opus 5.5’s in the evaluated settings. That is a benchmark-based estimate, not a guarantee that K3.1 Pro will be cheaper for all workloads. And self-hosting would not automatically be inexpensive: the source report notes that a model with 2.8 trillion total parameters would require substantial computing infrastructure if run independently.

The strongest reported improvement is in agentic testing. Artificial Analysis measured 67% on Terminal-Bench 4.0 for K3.1 Pro, compared with 13% for the earlier K3. On the GDPval-AA knowledge-work evaluation, the model scored 1,846 Elo, behind Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. Those results suggest gains in tested tasks, but do not establish performance across all coding or office work.

From K3 to K3.1 Pro

K3.1 Pro is described in the source report as a more heavily post-trained version of Kimi K3. The earlier K3 scored 43.6 on Artificial Analysis’s Intelligence Index in the comparison. On the same report’s account, K3.1 Pro also improved on Humanity’s Last Exam, moving from 47% to 52%, while Opus 5.5 scored 61%. The benchmarks use different tasks and should be read as measures of performance on those evaluations, not as a single verdict on general capability.

Moonshot’s launch material also presents demonstrations in game development, 3D modeling and finance. The company says the model can build projects using Unity, Unreal Engine and Godot, and can write scripts for Blender and CAD tools. On BenchCAD with tools, Moonshot reports a score of 92.7%, compared with 96.2% for Opus 5.5 and 95.9% for GPT-6 Astra. These are vendor-reported results; the demonstrations were not independently verified in the cited report.

Moonshot also advertises a roughly 65% reduction in hallucinations from K3’s 23.4% to K3.1 Pro’s 8.3% on its internal Kimi Hallu Bench Hard test. The source says Moonshot describes that test as deliberately designed to provoke errors and cautions that its results do not represent everyday error rates. Artificial Analysis measured a smaller change on its separate AA-Omniscience evaluation: 53% for K3 and 44% for K3.1 Pro.

“The test is deliberately crafted to provoke errors.”

— Moonshot AI

Weights and Testing Still Pending

The main unresolved issue is whether Moonshot will release the model weights by its stated Nov. 4 target, and what licence will govern their use. Without the weights and terms, users cannot assess whether the model will be available for self-hosting or what restrictions might apply. Artificial Analysis’s open-weight comparison is explicitly dependent on that release.

There are also limits to what the current numbers establish. Artificial Analysis provides independent benchmark results, but benchmark scores do not settle how a model will behave in a live deployment, and Moonshot’s claims about games, CAD, finance and its internal hallucination test remain company-reported in the source material. The report does not provide full details of the independent test methodology beyond the listed evaluation settings and scores.

The source reports K3.1 Pro at $2 per million input tokens and $10 per million output tokens, but it does not establish whether those prices will remain unchanged. It also does not provide the hardware requirements, operating costs or practical performance figures for running the model locally. Those details will matter to organizations considering deployment, particularly given its reported total parameter count.

Watch for the Nov. 4 Release

The next major milestone is Moonshot’s promised release of K3.1 Pro’s weights by Nov. 4. A release would allow users and outside evaluators to inspect and test the model more directly; the accompanying licence will determine how it can be used, modified and redistributed. Until then, claims that it is an available open-weight alternative remain prospective rather than a description of what users can download today.

Further independent evaluations could clarify whether the launch-day benchmark results hold across additional tasks and real-world deployments. For now, the confirmed picture is narrower: K3.1 Pro is available through Moonshot’s products and API, has scored strongly in Artificial Analysis’s reported testing, and remains short of an independently usable open-weight release. The promised weights, licence and follow-up testing will determine how much of its benchmark showing translates into a practical option for users and organizations.

Key Questions

What is Kimi K3.1 Pro?

Kimi K3.1 Pro is a Moonshot AI model described as a more heavily post-trained version of Kimi K3. Artificial Analysis lists 2.8 trillion total parameters, 104 billion active parameters, text and image input, and a one-million-token context window.

How did K3.1 Pro rank?

It scored 55.1 on Artificial Analysis’s Intelligence Index in the reported comparison, placing fourth. The listed leader, Claude Opus 5.5, scored 57.6, a difference of 2.5 points.

Can users download the model weights now?

No weight release is reported yet. Moonshot has promised the weights by Nov. 4, and the source report says the licence has not been published. The model is available through Kimi products and the API.

Are Moonshot’s hallucination claims independently confirmed?

Not in the cited source. Moonshot reports a decrease from 23.4% to 8.3% on its internal test, while Artificial Analysis measured a change from 53% to 44% on a different evaluation. The results use different tests and are not directly interchangeable.

How much does API use cost?

The listed price is $2 per million input tokens and $10 per million output tokens. Artificial Analysis estimated $2.11 per Intelligence Index task under its evaluation, but actual costs will depend on usage and the applicable pricing.

Source: ThorstenMeyerAI.com