Moonshot launches Kimi K3.1 Pro, 4th in AA Index, price drop

Key Takeaways

1Kimi K3.1 Pro scores 55 points and ranks 4th on the Artificial Analysis Intelligence Index

2Prices drop by up to 33%, with a cost per task of $1.98

3Moonshot removes cybersecurity data for weight publication, measuring 81.0% on xdev-bench-hard after cleaning

4The model progresses on agentic, office, and 3D tasks, closing the gap with top proprietary models

Why it matters : K3.1 Pro could become the highest-rated open-weight model, narrowing the gap with proprietary leaders while framing its offensive capabilities.

Moonshot AI launches the Kimi K3.1 Pro and makes firm choices in cybersecurity ahead of a planned weight release on November 4, 2026. The model shows significant progress in agentic tasks, programming, and 3D, climbing to 55 points on the Artificial Analysis Intelligence Index and seeing its prices drop by up to 33%. Internal and public comparisons now place the K3.1 Pro just a few points behind the best proprietary models.

Moonshot Addresses Cybersecurity and Cleans Sensitive Data

Moonshot AI describes an increase in vulnerability research capabilities as post-training intensity rises, due to the presence of cybersecurity content in the intermediate corpus and associated tasks in the RL environment. The company specifies that it excluded cybersecurity-related information during both training phases to allow for an open-access weight release. On its internal xdev-bench-hard, the version that was not purged achieves a 90.5% success rate on tasks, while after the removal of sensitive data, the K3.1 Pro solves 81.0%. In the same test, DeepSeek-V4.1-Flash scores 85.7%, while Kimi K3 achieved 33.3%. Moonshot asserts that no open-weight model should exceed the ultimate limit of offensive capabilities, while gradually seeking to develop defensive uses, such as auditing legacy systems. Anthropic has also demonstrated that GLM-5.3, rated 45 on the AA Index, can write exploits at levels close to Claude Mythos.

Prices Drop and Cost per Task at $1.98

Moonshot reduces its prices by 33%. The entry price drops to $2 per million tokens from $3 previously, the output to $10 from $15, and cache hits to $0.20 from $0.30. The cost per task comes out to $1.98 for K3.1 Pro, compared to $2.00 for K3, a level also observed for each task on the Intelligence Index, in a context where the model generates 14% more output tokens. In terms of size, K3.1 Pro retains a total of 2.8 trillion parameters, of which 104 billion are active, with a context window of one million tokens and support for both text and image. Artificial Analysis places it on the Pareto frontier and indicates that no model of comparable size performs better on its Index, positioning it on its Pareto line.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Ranking: 55 Points and 4th Place, 3 Points from the Top

Kimi K3.1 Pro reaches 55 points and 4th place on the Artificial Analysis Intelligence Index, just behind Claude Opus 5.5. The gap with the top trio—Claude Opus 5.5 at 58 points, Claude Sonnet 5.5, and Muse Wonder Full at 56—narrows to 3 points compared to proprietary models, with an 11-point gain over K3. Artificial Analysis estimates that if Moonshot releases the weights on November 4, 2026, as planned, K3.1 Pro would become the highest-rated open-weight model, and the gap between open and proprietary would shrink from 12 to 3 points, assuming no other changes occur. Moonshot aims for this weight release by November 2026.

Agentic, Terminal, and Hallucinations: Strong Progress on Benches

On GDPval-AA, K3.1 Pro shows 1,846 Elo, which is 312 points better than K3, behind Claude Opus 5.5 at 1,866 and Muse Wonder Full at 1,857. On Terminal-Bench 4.0, Moonshot reports 65.7% for K3.1 Pro, above GPT-6 Astra at 58.2% and close to Claude Opus 5.5 at 66.4%, while K3 only reached 13%. Hallucinations at the claim level decrease on the internal Kimi Hallu Bench Hard, from 23.4% with K3 to 8.3% with K3.1 Pro; Moonshot specifies that this is a stress test. On the same bench, GPT-6 Astra is measured at 5.5%. On internal financial benches, K3.1 Pro gains 7.2 points in modeling and 11.6 points in research, outperforming all tested models on OfficeQA Pro v2 according to the company.

3D and Game Engines: BenchCAD at 92.7% and 3D Make Tool

Moonshot indicates that K3.1 Pro interacts directly with Unity, Unreal Engine, and Godot, for quick fixes as well as project builds. On BenchCAD with tools, the model achieves 92.7%, close to Claude Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%; without tools, it scores 72.8%, nearly on par with Claude Opus 5.5 at 73.0%. The company highlights notable performance in parametric modeling of mechanical components, architectural structures, and customizable products. The model is capable of generating and executing scripts for Blender as well as CAD software without relying on MCP connectors. Moonshot is concurrently rolling out the 3D Make function on Kimi.com to preview, edit, and export 3D models.