Three Points Short: Kimi K3.1 Pro’s Frontier-Level Ambition

🔍 Read the full analysis: Three Points Short: Kimi K3.1 Pro’s Frontier-Level Ambition on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get smart everyday buys delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Moonshot AI launched Kimi K3.1 Pro on October 23, with the model available through its apps, Kimi Work, Kimi Code and API. Artificial Analysis scored it 55.1 on its Intelligence Index, placing it fourth in the comparison—three points below Claude Opus 5.5—and listed input and output prices of $2 and $10 per million tokens, respectively. The weights have not yet been released; Moonshot says they are due by November 4.

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 listed one-million-token context window. The benchmark figures cited here are from Artificial Analysis’s model page as of October 24, with models evaluated at their highest reasoning setting where available.

In that comparison, K3.1 Pro’s score of 55.1 trails Opus 5.5 at 57.6, Sonnet 5.5 at 56.0 and Muse Wonder Full at 55.8. It scored 67% on Terminal-Bench 4.0, the highest result in the cited table, and reached 1,846 Elo on GDPval-AA, behind Opus 5.5 and Muse Wonder Full. On Humanity’s Last Exam it scored 52%, compared with 61% for Opus 5.5.

Artificial Analysis estimates K3.1 Pro costs $2.11 per Intelligence Index task, against $5.98 for Opus 5.5 at maximum effort. Those figures reflect the test’s task costs, not a guaranteed price for every use. The listed API rates are $2 per million input tokens and $10 per million output tokens. Moonshot’s performance demonstrations and internal-test claims are separate from these independent benchmark results.

A Narrower Gap, With Conditions

The score places K3.1 Pro near the top of a comparison dominated by proprietary models. If Moonshot releases usable weights under terms that allow the expected forms of access, developers may be able to run or adapt a model with a benchmark score close to the leading hosted systems. That could matter to organisations weighing performance against API costs, data handling and control over deployment.

For now, that possibility is conditional. The model is not yet available as published weights, and a licence for K3.1 has not been announced in the source material. Artificial Analysis describes it as set to become the leading open-weights model once the weights are published; that is a forecast dependent on the release, not a description of what users can download today. The listed scale—2.8 trillion total parameters—also means self-hosting would require substantial infrastructure rather than ordinary consumer hardware.

The results suggest a particularly sharp change in the tested agentic tasks: Terminal-Bench rose from 13% for Kimi K3 to 67% for K3.1 Pro in Artificial Analysis’s evaluation. Other gains were more limited. The benchmark spread matters to readers because a high aggregate score does not mean equal performance across coding, knowledge work and factual accuracy.

From Kimi K3 to K3.1 Pro

Moonshot describes K3.1 Pro as a more heavily post-trained version of Kimi K3. The distinction helps explain why the launch focuses on benchmark and task performance rather than presenting a wholly new model architecture. In Artificial Analysis’s cited comparison, the earlier K3 scored 43.6 on the Intelligence Index, while K3.1 Pro scored 55.1.

Before this launch, the highest-scoring open-weights model listed in the supplied comparison was Xiaomi’s MiMo-V2.6-Pro at 46.3. K3.1 Pro’s higher score is meaningful as a benchmark result, but its status as an open-weights model depends on the promised release. Moonshot’s product announcement and Artificial Analysis’s evaluations also answer different questions: the first describes the company’s claimed capabilities, while the latter provides third-party benchmark measurements.

Moonshot’s launch materials highlight projects in Unity, Unreal Engine and Godot, parametric modelling in Blender and CAD, and finance-related work. The company reports a 92.7% result on BenchCAD with tools, compared with 96.2% for Opus 5.5 and 95.9% for GPT-6 Astra. These are vendor-reported results and should not be treated as independent confirmation of the demonstrations or their performance in routine workplace use.

“Kimi K3.1 Pro is its most capable “3T-class open-weight model.””

— Moonshot AI

Weights, Licence and Real-World Accuracy

As of October 24, K3.1 Pro’s weights were not public, and the source material gives no published licence for the model. It is therefore unclear whether the November 4 target will be met, what use or redistribution the licence will allow, and how accessible the model will be to developers outside Moonshot’s hosted services.

There is also a gap between the company’s internal claims and independent measurements. Moonshot says its Kimi Hallu Bench Hard test showed claim-level hallucinations falling from 23.4% for K3 to 8.3% for K3.1 Pro, a reduction of about 65% on that test. Artificial Analysis measured a decrease from 53% to 44% on AA-Omniscience, a smaller change on a different evaluation. These results cannot be treated as a direct apples-to-apples comparison, and neither establishes error rates across everyday use.

The supplied material does not establish how the model performs across a broad range of users, languages or production workloads. Moonshot’s examples of game development, CAD work and finance tasks are company claims; independent replication and practical deployment details remain limited.

The November 4 Weight Release

The next key milestone is Moonshot’s stated plan to publish K3.1 Pro’s weights by November 4. A release would allow researchers and developers to inspect the model and test whether its benchmark performance carries over to local or independently managed deployments. The accompanying licence will be central to determining what those users can do with it.

Readers should also watch for further independent evaluations, particularly on the vendor’s claimed software-building, CAD and finance workflows, and for clearer details about the compute needed to run the model. Until the weights, licence and those results are available, K3.1 Pro’s benchmark standing is measurable, but its practical status as an open model remains unresolved.

Key Questions

What is Kimi K3.1 Pro?

It is a model launched by Moonshot AI on October 23, 2026. Artificial Analysis lists 2.8 trillion total parameters, 104 billion active parameters, a one-million-token context window, and text and image input.

How did it rank against other models?

Artificial Analysis gave it 55.1 on the Intelligence Index, fourth in the comparison provided as of October 24. It trailed Claude Opus 5.5 by three points.

Can developers download the model now?

No. The source material says the weights were promised by November 4 but had not been released by October 24. K3.1’s licence had also not been published.

How much does Kimi K3.1 Pro cost?

The listed API price is $2 per million input tokens and $10 per million output tokens. Artificial Analysis estimates $2.11 per Intelligence Index task; actual costs can vary by use.

Are Moonshot’s hallucination claims independently confirmed?

Not in the same test. Moonshot reported a reduction from 23.4% to 8.3% on its internal Kimi Hallu Bench Hard test, while Artificial Analysis measured a change from 53% to 44% on AA-Omniscience. The figures come from different evaluations and do not establish routine-use error rates.

Source: ThorstenMeyerAI.com