Kimi K3.1 Pro’s AI Score Nears The Frontier, While Weights Remain Out

🔍 Read the full analysis: Kimi K3.1 Pro’s AI Score Nears The Frontier, While Weights Remain Out 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 launched Kimi K3.1 Pro on Oct. 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, Kimi Work, Kimi Code and API, but its weights have not been released; the company says they are due by Nov. 4, and has not published a licence for them.

Artificial Analysis’s ranking puts K3.1 Pro behind Claude Opus 5.5 at 57.6, Claude Sonnet 5.5 at 56.0 and Muse Wonder Full at 55.8. The assessment compares models at their highest reasoning settings. Its figures are a snapshot of the model page as of Oct. 24, not a guarantee of performance across every task or use case. The score puts K3.1 Pro 3.5 points behind the leader in that evaluation.

Artificial Analysis lists the model as a mixture-of-experts system with 2.8 trillion total parameters and 104 billion active parameters, a one-million-token context window, and text and image input. The listed API prices are $2 per million input tokens and $10 per million output tokens. Artificial Analysis estimates a cost of $2.11 per Intelligence Index task, compared with $5.98 for Claude Opus 5.5 at maximum effort. Those figures are tied to the evaluator’s task and pricing assumptions.

The independent evaluation shows a marked result on agentic coding: K3.1 Pro scored 67% on Terminal-Bench 4.0, compared with 13% for Kimi K3, its predecessor. On GDPval-AA, a test of simulated knowledge work, it recorded 1,846 Elo, behind Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. Its Humanity’s Last Exam score was 52%, while Opus 5.5 scored 61%. These results indicate strengths in particular evaluations, not universal superiority.

A Near-Frontier Score, Without the Weights

The gap between a strong proprietary model and a model whose weights can be downloaded affects who can deploy AI, where data is processed and how much control organisations have over updates and infrastructure. Artificial Analysis’s score suggests K3.1 Pro could narrow that performance gap if Moonshot releases the weights as promised. But the model is not yet available to run independently, so its current API and app access should not be confused with an open-weight release.

Access to weights alone would not make deployment simple. Artificial Analysis lists 2.8 trillion total parameters, making self-hosting a substantial infrastructure undertaking. The eventual licence will also determine how the weights can be used, including whether commercial deployments are permitted and what other conditions apply. Until Moonshot publishes it, businesses cannot judge those terms from the launch announcement.

The reported prices may also draw interest from teams comparing hosted models. Artificial Analysis’s per-task estimate is lower than its figure for Opus 5.5, but it depends on the evaluator’s benchmark workload and provider rates. It is not a prediction of a company’s operating costs, which can vary with prompt length, usage patterns and implementation.

How K3.1 Compares With Kimi K3

K3.1 Pro is described as a more heavily post-trained version of Kimi K3. In the Artificial Analysis results supplied for this report, the earlier K3 scored 43.6 on the Intelligence Index, compared with 55.1 for K3.1 Pro. The predecessor is listed as open under the Kimi K3 License, while K3.1 Pro is shown as having weights promised by Nov. 4 rather than weights already available.

Moonshot’s launch materials also promote software-building, 3D modelling, finance and reduced hallucinations. The company says the model can produce projects using Unity, Unreal Engine and Godot, and can write scripts for Blender and CAD software. Those are vendor claims based largely on its own demonstrations and tests. For example, Moonshot reports a 92.7% result on BenchCAD with tools, below the reported 96.2% for Opus 5.5 and 95.9% for GPT-6 Astra.

Moonshot also reports that 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. The company describes the test as deliberately designed to provoke errors and says its rates do not represent overall error rates in daily use. Artificial Analysis recorded a reduction from 53% to 44% on its separate AA-Omniscience evaluation, a smaller change under that benchmark. The two measures use different tests and should not be treated as interchangeable.

“The Kimi Hallu Bench Hard test is “deliberately crafted” to provoke errors, and the rates “do not represent overall error rates in daily use.””

— Moonshot AI

Weights, Licence and Real-World Results

The main unresolved issue is whether Moonshot will publish the weights by Nov. 4, as promised. The source material does not provide a published licence, so the permitted uses and conditions remain unknown. Artificial Analysis’s description of K3.1 Pro as a leading open-weight model is explicitly conditional on publication.

There are also limits to what the current evaluations establish. The Intelligence Index is one independent comparison, while many of Moonshot’s claims about full software projects, finance performance and hallucinations rely on company-run tests or demonstrations. The available information does not show how those results will hold across varied production tasks, or how the model will perform when deployed outside Moonshot’s services.

Self-hosting requirements, hardware configurations and costs have not been detailed in the supplied material. The reported parameter count points to a large system, but does not by itself establish what configuration will be needed to run it. Independent testing of the released weights will be needed to assess reproducibility and practical deployment conditions.

The Nov. 4 Release Is the Test

The next stated milestone is Moonshot’s promised Nov. 4 weights release. If it arrives, the accompanying licence and release details will determine whether developers and organisations can download, modify and deploy the model, and under what terms. The release could also allow independent researchers to check the model against the results reported before the weights were available.

Until then, users can access K3.1 Pro through Moonshot’s apps, Kimi Work, Kimi Code and API, subject to the provider’s service terms and listed prices. The ranking may change as evaluation pages are updated or new models are tested. For now, the confirmed picture is a high score on one independent index and a pending release of the materials needed to verify open-weight access.

Key Questions

What did Artificial Analysis score Kimi K3.1 Pro?

Artificial Analysis listed K3.1 Pro at 55.1 points on its Intelligence Index as of Oct. 24, 2026, placing it fourth in the comparison supplied. The ranking reflects particular tests and settings, not every task.

Can developers download K3.1 Pro now?

No. Moonshot made the model available in its apps, Kimi Work, Kimi Code and API, but the weights have not been published. The company said they are due by Nov. 4.

Has Moonshot published a licence for the weights?

No licence for K3.1 Pro was published in the source material. Until Moonshot releases the terms, permitted uses and conditions are unclear.

How much does Kimi K3.1 Pro cost through the API?

The listed price is $2 per million input tokens and $10 per million output tokens. Artificial Analysis estimated $2.11 per Intelligence Index task; that benchmark-specific estimate is not a general operating-cost forecast.

Are Moonshot’s performance and hallucination claims independently confirmed?

Not all of them. Artificial Analysis independently evaluated the model on its own benchmarks, while claims about specific demos and Moonshot’s internal Hallu Bench come from the company. On a separate Artificial Analysis hallucination test, the reported rate fell from 53% to 44%; Moonshot’s internal figures use a different test.

Source: ThorstenMeyerAI.com