Kimi K3.1 Pro Delivers A Near-Frontier AI Score

🔍 Read the full analysis: Kimi K3.1 Pro Delivers A Near-Frontier AI Score 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, and Artificial Analysis scored it 55.1 on its Intelligence Index, placing it fourth in the comparison reported the next day. The result puts the model close to the highest-scoring proprietary systems in that test, but its weights are not yet available: Moonshot says they are due by November 4, and 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. The listing gives it a 1-million-token context window and text and image input. Moonshot says the model is a more heavily post-trained version of Kimi K3. It went live in the Kimi apps, Kimi Work, Kimi Code and the Kimi API on October 23.

In Artificial Analysis’s comparison, K3.1 Pro’s 55.1 score trails Claude Opus 5.5 at 57.6, Claude Sonnet 5.5 at 56.0 and Muse Wonder Full at 55.8. The evaluator gives K3.1 Pro an estimated $2.11 cost per Intelligence Index task, compared with $5.98 for Opus 5.5 at its highest reasoning setting. These figures come from a specific benchmark and pricing comparison; they do not establish that the models perform identically across all tasks.

The evaluation also found a pronounced gap on Terminal-Bench 4.0, a test of agentic coding tasks: K3.1 Pro scored 67%, compared with 13% for Kimi K3. Artificial Analysis reported 1,846 Elo on GDPval-AA, its simulated knowledge-work evaluation, behind Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. On Humanity’s Last Exam, K3.1 Pro scored 52%, while Opus 5.5 scored 61%.

A Narrower Gap in Benchmark Scores

The result matters because open-weight models have generally scored below leading closed systems in the comparison cited by the source. The day before K3.1 Pro’s launch, the top open-weight model listed by Artificial Analysis was Xiaomi’s MiMo-V2.6-Pro at 46.3. K3.1 Pro’s 55.1 score would substantially narrow that measured gap if its weights are released as promised.

That distinction is not just about benchmark rankings. Weights that can be downloaded may give organisations more control over deployment, including the option to run a model on their own infrastructure. But that possibility depends on the actual release and its licence. The model’s reported scale—2.8 trillion total parameters—also means that self-hosting could require substantial computing resources, rather than being a practical option for every developer or company.

The reported price comparison may also draw attention from teams choosing models for repeated workloads. Artificial Analysis estimates a lower cost per index task for K3.1 Pro than for Opus 5.5 in its test. That is a benchmark-specific estimate, not a guarantee of lower costs for a particular application: actual expenses and performance can vary with prompts, usage, reasoning settings and infrastructure.

What the Benchmark Measures

The Intelligence Index is an Artificial Analysis evaluation, not a direct measure of every model capability or a guarantee of performance in everyday use. The figures reported here compare models at their listed highest reasoning settings, with Muse Wonder Full shown at its maximum setting and Gemini 4 Argon at high. Rankings can change as evaluators update models or methods.

Moonshot has separately promoted results from its own demonstrations and internal tests. It says K3.1 Pro can complete projects involving Unity, Unreal Engine and Godot, and can produce parametric 3D work in Blender and CAD software. The company also reports an internal finance-modeling improvement over K3 and a 64.4 result on OfficeQA Pro v2. These are vendor-reported claims, and the source material says the demonstrations have not been independently verified.

Moonshot’s headline claim of 65% fewer hallucinations comes from its internal Kimi Hallu Bench Hard test, where the reported claim-level hallucination rate fell from 23.4% for K3 to 8.3% for K3. Artificial Analysis measured a smaller reduction on a different evaluation: AA-Omniscience recorded a rate falling from 53% to 44%. Moonshot says its internal test was deliberately designed to provoke errors and does not represent overall error rates in daily use. The results should not be treated as interchangeable measures.

Weights, Licence and Real-World Results

The weights had not been released in the reporting available on October 24. Moonshot’s November 4 target is a stated plan, not confirmation that publication has occurred. The company also had not published a K3.1 licence, leaving open what uses, modifications or redistribution the eventual terms will allow.

There are further limits to the available evidence. Artificial Analysis has published benchmark results, but those scores cannot establish how the model will perform across every coding, business or research task. Moonshot’s game-building, CAD and finance examples are company claims; the supplied material does not include independent verification of those demonstrations. The source also ends partway through its discussion of the AA-Omniscience comparison, so no further result from that section can be confirmed here.

The model’s cost and hardware requirements also depend on how it is accessed. The API prices cited are $2 per million input tokens and $10 per million output tokens. Those hosted-service rates do not tell readers what it would cost to run the full model themselves, and the reported parameter count suggests self-hosting may demand data-centre-scale resources.

Watch for the November Weights Release

The next concrete milestone is Moonshot’s promised November 4 weight release. Its arrival, the accompanying licence and the files actually made available will determine whether K3.1 Pro can be independently inspected and deployed beyond Moonshot’s services. If the date or terms change, the current conditional claim about its standing among open-weight models will need to be revisited.

In the meantime, Artificial Analysis’s published scores and Moonshot’s API offering provide the available basis for comparison. More independent testing could clarify how the benchmark gains translate to particular workloads, especially agentic coding and knowledge work. For companies considering adoption, the open questions are the model’s performance on their own tasks, the total cost of use and whether the eventual release terms fit their deployment needs.

Key Questions

How did Kimi K3.1 Pro rank on Artificial Analysis’s Intelligence Index?

It scored 55.1 and ranked fourth in the comparison reported on October 24, behind Claude Opus 5.5 at 57.6, Claude Sonnet 5.5 at 56.0 and Muse Wonder Full at 55.8.

Is Kimi K3.1 Pro open-weight now?

No. Moonshot said the weights were due by November 4, but the supplied reporting says they had not been released as of October 24. Artificial Analysis described the model’s prospective open-weight standing conditionally on publication.

What does Moonshot mean by 65% fewer hallucinations?

The figure comes from Moonshot’s internal Kimi Hallu Bench Hard test, where it reported a claim-level rate falling from 23.4% for K3 to 8.3% for K3.1 Pro. Artificial Analysis recorded a different, smaller reduction on AA-Omniscience, from 53% to 44%; the tests measure different things.

How much does Kimi K3.1 Pro cost to use?

The listed API price is $2 per million input tokens and $10 per million output tokens. Artificial Analysis estimated $2.11 per Intelligence Index task in its comparison. That task estimate is not a universal price for every use case.

Can companies run the model on their own servers?

That is not yet established. The weights were still pending and Moonshot had not published a licence as of October 24. The model’s listed 2.8 trillion total parameters also suggest that self-hosting could require substantial infrastructure.

Source: ThorstenMeyerAI.com