The Frontier Is Three Points Away For Kimi K3.1 Pro

🔍 Read the full analysis: The Frontier Is Three Points Away For Kimi K3.1 Pro on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

• 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 released Kimi K3.1 Pro in its apps, Kimi Work, Kimi Code and API on October 23, and an Artificial Analysis evaluation gave it an Intelligence Index score of 55.1. That put it fourth among the models listed, three points behind Claude Opus 5.5; Moonshot says the model’s weights will be published by November 4, but no K3.1 licence has been published.

Artificial Analysis’s October 24 model-page figures rank 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 evaluation lists K3.1 Pro at 2.8 trillion total parameters and 104 billion active parameters, using a mixture-of-experts design. It accepts text and images and has a stated context window of one million tokens.

The same evaluator lists API pricing at $2 per million input tokens and $10 per million output tokens, and estimates $2.11 per Intelligence Index task. Its estimate for Opus 5.5 is $5.98 per task under the listed evaluation conditions. These are benchmark-based cost estimates, not a guarantee of what every real-world workload will cost.

Artificial Analysis measured a sharp increase on its Terminal-Bench 4.0 agentic coding test: 67% for K3.1 Pro, compared with 13% for Kimi K3. The new model scored 1,846 Elo on GDPval-AA, behind Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. On Humanity’s Last Exam, it scored 52%, while Opus 5.5 scored 61%.

A Smaller Gap to Closed Models

If Moonshot publishes usable weights as promised, the benchmark results could make K3.1 Pro a notable option for organizations seeking a model they can run or adapt themselves. Artificial Analysis describes it conditionally as set to become the leading open-weights model once the weights are published. The qualifier matters: at present, customers can access the model through Moonshot’s services, but the reported weights are not yet available.

The results also suggest that performance varies by task. K3.1 Pro’s strongest measured showing in the supplied figures is agentic coding, while its knowledge-test score remains below Opus 5.5. Companies comparing models will need to weigh task performance, access arrangements, licensing and infrastructure costs—not just a single overall ranking. Running a model described as having 2.8 trillion total parameters would also require substantial computing resources; open weights would not, by themselves, make local deployment practical for every user.

Price comparisons should be read in their stated context. Artificial Analysis’s per-task figure reflects its benchmark setup, while actual API bills depend on usage and task design. The listed token rates are comparable to some competing APIs, but the evaluator’s estimated cost per Index task is lower than its estimate for Opus 5.5.

What the Evaluations Measure

Moonshot presents K3.1 Pro as a more heavily post-trained version of Kimi K3. Its launch materials describe work across coding, software agents and knowledge tasks. Artificial Analysis independently evaluated the model and compared it with other systems at their highest listed reasoning settings, but its results are still benchmark measurements rather than proof of performance in every setting.

Moonshot’s launch post also describes projects built with game engines including Unreal, Unity and Godot, and parametric 3D work in Blender and CAD. The company reports a 92.7% result on BenchCAD with tools, below the figures it gives for Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%). These examples and several other results are Moonshot’s own demonstrations or internal tests, not independent verification of routine performance.

The difference is clear in the hallucination figures. Moonshot says its internal Kimi Hallu Bench Hard test showed claim-level hallucinations falling from 23.4% for Kimi K3 to 8.3% for K3.1 Pro, a reduction of about 65%. Artificial Analysis’s AA-Omniscience evaluation recorded a smaller change, from 53% to 44%. Moonshot says its test is deliberately designed to provoke errors and does not represent overall daily-use error rates.

“The test is deliberately crafted to provoke errors.”

— Moonshot AI

Weights and Licensing Remain Pending

The most consequential open question is whether Moonshot will publish the weights by its stated November 4 target. Until they are released, users cannot independently inspect or run the model from public weights, and Artificial Analysis’s description of K3.1 Pro as an open-weights leader remains prospective. Moonshot has not published a K3.1 licence, leaving unclear what uses, modifications or redistribution terms would be allowed.

It is also not clear how well the launch demos and benchmark gains will carry over to ordinary commercial work. Moonshot’s game-development, CAD and finance examples are vendor claims, and the supplied material does not provide independent replication of those demonstrations. Benchmark scores can help compare models under defined tests, but they do not establish results across every task or workflow.

The supplied figures list prices and scores but do not resolve the full cost of deploying a model of this scale on a company’s own infrastructure. Hardware requirements, serving efficiency and practical operating costs will depend on implementation and workload. Those details matter if customers are deciding between Moonshot’s hosted API and a self-managed deployment.

The November 4 Weights Deadline

The next stated milestone is Moonshot’s promised release of K3.1 Pro’s weights by November 4. Publication would allow independent users to examine the files and test whether the model can be run under their own conditions; the licensing terms will be central to determining whether commercial use and modification are permitted.

Until then, the model is available through the Kimi apps, Kimi Work, Kimi Code and the Kimi API. Readers should distinguish those current hosted-access options from the planned weights release. Further independent evaluations, especially on real-world coding, knowledge work and reliability, could clarify how much of the reported performance translates beyond benchmark tasks.

Key Questions

What is Kimi K3.1 Pro?

Kimi K3.1 Pro is Moonshot AI’s updated model, released in its apps, Kimi Work, Kimi Code and API on October 23, 2026. Artificial Analysis lists 2.8 trillion total parameters, 104 billion active parameters and a one-million-token context window.

How did it rank against other models?

Artificial Analysis gave it an Intelligence Index score of 55.1 and ranked it fourth in the listed comparison, behind Claude Opus 5.5, Claude Sonnet 5.5 and Muse Wonder Full. The score is a benchmark result, not a guarantee of performance on every task.

Are the model weights available now?

No. Moonshot says the weights are expected by November 4. Until they are published, the claimed open-weight status and the evaluator’s prospective open-model ranking have not been realized.

What licence will apply to the weights?

No K3.1 licence has been published in the supplied source material. The terms for commercial use, modification and redistribution are therefore unclear.

Are Moonshot’s performance claims independently verified?

Some performance figures come from Artificial Analysis’s independent evaluations, including the Intelligence Index and Terminal-Bench results. Claims about demos, internal finance tests and Moonshot’s hallucination benchmark come from the company and have not been independently verified in the source material.

Source: ThorstenMeyerAI.com