Kimi K3.1 Pro approaches top proprietary top models

Moonshot AI has released Kimi K3.1 Pro, a post-trained version of its reasoning model. It climbs to fourth place in the Artificial Analysis Intelligence Index and reduces its API prices by 33%.

Moonshot AI has launched Kimi K3.1 Pro, a post-trained version of its Kimi K3 model, with marked gains in agentic reasoning, coding, and 3D modeling. According to Artificial Analysis, the model scores 55 points on the Intelligence Index, moving up four high-level ranks and narrowing the gap to 3 points with the best proprietary models. At this stage, the challenge goes beyond mere ranking: the Chinese company Moonshot narrows the distance between open and closed models while lowering its prices.

Un bond net sur le raisonnement et les tâches d’agent

The most visible progress concerns so-called agentic tasks, that is, scenarios where the model chains actions, tools, and decisions to carry out a task from start to finish. On GDPval-AA, Kimi K3.1 Pro reaches 1,846 Elo, compared to 1,534 for K3 [to be verified], placing just behind Claude Opus 5.5 at 1,866. On Terminal-Bench 4.0, Moonshot reports 65.7%, nearly at the level of Claude Opus 5.5 (66.4%) and ahead of GPT-6 Astra (58.2%). The model changes less discourse than muscle, but the result is visible in the score tables.

Moonshot also reports a sharp drop in hallucinations on its internal test Kimi Hallu Bench Hard: 23.4% with K3, versus 8.3% with K3.1 Pro. The company specifies this is a stress test and not a general error measure. On financial research benchmarks, the model gains 7.2 points in Financial Modeling and 11.6 points in Financial Research, while taking the lead on OfficeQA Pro v2 for office tasks, according to the company.

Le code, Unity et Blender entrent dans la boucle

For development uses, Kimi K3.1 Pro approaches the strongest models on the market. Moonshot says the system works directly with Unity, Unreal Engine, and Godot, can quickly correct code as well as produce complete projects, and can write and run scripts for Blender and CAD software without an MCP connector. On BenchCAD with tools, it scores 92.7%, compared to 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra. Without tools, it reaches 72.8%, nearly at the level of Claude Opus 5.5 at 73.0%.

The 3D aspect is not trivial. Moonshot highlights work on parametric mechanical parts, architectural models, and configurable products, with a “3D Make” feature on Kimi.com to preview, modify, and export models. The model also maintains a context window of one million tokens, a format that allows processing long documents or codebases without losing track too quickly.

Des prix plus bas, mais le même gabarit

Kimi K3.1 Pro retains the same architecture as K3, with 2.8 trillion parameters in total and 104 billion active. The difference lies in post-training, thus in how this computational mass is exploited rather than the raw size of the model. According to Artificial Analysis, it sits on the Pareto frontier, meaning the point where it becomes difficult to significantly improve one criterion without degrading another.

Moonshot simultaneously lowers its API prices by 33%. Entry costs $2 per million tokens, down from $3; output is $10, down from $15; and cache costs $0.20, down from $0.30. The cost per task on the Intelligence Index comes to $1.98, almost the same level as K3 at $2.00, because the model generates more tokens. Performance promises are one thing, but the bill often decides production deployment.

Cyber et open weights : Moonshot met des limites

Moonshot’s technical post also devotes a large section to cybersecurity. The company explains it removed cyber-related data from training phases for the open-weights release, as it observed that vulnerability research capabilities increased with the scale of post-training. On xdev-bench-hard, which reproduces recently disclosed vulnerabilities in the Linux kernel, Node.js, and JavaScriptCore, the uncleaned version solves 90.5% of tasks, versus 81.0% after data cleaning. Kimi K3 only reached 33.3% on this test.

Moonshot says it wants to avoid an open model pushing offensive capabilities too far, while allowing defensive uses such as auditing legacy systems. The company plans to release the model weights by November 2026. If this schedule holds, Kimi K3.1 Pro would become the best open weights model at the time according to the cited ranking. [to be verified]

Key points

  • Kimi K3.1 Pro scores 55 points on the Artificial Analysis Intelligence Index.
  • Moonshot announces a 33% reduction in API prices.
  • The model reaches 1,846 Elo on GDPval-AA.
  • The weights are announced for November 2026.
  • Moonshot AI is the actor cited in the sources.

By the numbers

  • 2.8 trillion parameters — total size of the Kimi K3.1 Pro model.
  • 104 billion active parameters — capacity actually mobilized at inference.
  • 1 million tokens — context window announced by Moonshot.
  • 33% — reduction in entry, output, and cache API prices.
  • 81.0% — K3.1 Pro score on xdev-bench-hard after cyber data cleaning.

To read