Moonshot’s Kimi K3.1 Pro narrows gap with Claude on AI benchmarks

Moonshot AI has released Kimi K3.1 Pro, a new reasoning model that The Decoder reports is now close to Anthropic’s leading Claude models on several benchmark rankings. The release matters because Moonshot says model weights are planned, which could make Kimi K3.1 Pro a stronger option for teams that want open-weight models for chat, coding and agent-style workflows.

A fourth-place score with open-weight ambitions

Kimi K3.1 Pro is a post-trained version of Moonshot’s earlier Kimi K3 model. According to Artificial Analysis, it scores 55 points on the Artificial Analysis Intelligence Index, an 11-point improvement over K3 and enough for fourth place.

The models ranked above it are Claude Opus 5.5 at 58 points, Claude Sonnet 5.5 at 56 points and Muse Wonder Full at 56 points, according to Artificial Analysis. That puts Kimi K3.1 Pro three points behind the top proprietary models cited in the report.

For readers comparing AI assistants and model providers, the most important part may be the planned weights release. The Decoder reports that if Moonshot releases Kimi K3.1 Pro’s weights by November 4, 2026 as planned, it would be the highest-scoring open-weights model available, assuming the ranking does not change before then.

Anthropic’s Claude remains a key reference point in the results, particularly Claude Opus 5.5. Kimi K3.1 Pro is not presented as clearly ahead across the board, but its position suggests open-weight models may be moving closer to top closed systems in tasks that matter to developers and advanced business users.

That context also fits a wider push around open and semi-open AI models, including recent activity covered by Emonarc such as Nvidia’s open-weight models for enterprise and PC agents and China’s draft rules for open-weight AI model exports.

Agentic tasks and office work improve sharply

The biggest gains appear in agentic tasks, where models are tested on multi-step work rather than single-turn answers. On GDPval-AA, Kimi K3.1 Pro reaches 1,846 Elo, according to Artificial Analysis. That is 312 points above K3 and just below Claude Opus 5.5 at 1,866 and Muse Wonder Full at 1,857.

Moonshot also reports a 65.7 percent score on Terminal-Bench 4.0. The Decoder notes this is ahead of GPT-6 Astra at 58.2 percent and just behind Claude Opus 5.5 at 66.4 percent. K3 scored 13 percent on the same benchmark.

For small teams using AI to handle research, spreadsheet-like analysis, internal documents or workflow automation, these benchmark shifts are worth watching. They do not prove that the model will perform better in every real deployment, but they suggest Moonshot has improved the model in areas closer to daily knowledge work.

Moonshot says Kimi K3.1 Pro also leads tested models on OfficeQA Pro v2. On internal finance benchmarks, the company reports gains of 7.2 points in Financial Modeling and 11.6 points in Financial Research.

The company also says claim-level hallucinations fell on its internal Kimi Hallu Bench Hard, from 23.4 percent with K3 to 8.3 percent with K3.1 Pro. Moonshot describes that benchmark as a stress test rather than a general error-rate measure. GPT-6 Astra scores 5.5 percent on the same test, according to the source, so it remains ahead there.

Coding, 3D modeling and pricing

Kimi K3.1 Pro also targets coding and 3D creation workflows. Moonshot says the model can work directly with Unity, Unreal Engine and Godot, including both fixes and full project builds.

On BenchCAD with tools, it scores 92.7 percent, compared with 96.2 percent for Claude Opus 5.5 and 95.9 percent for GPT-6 Astra. Without tools, Kimi K3.1 Pro reaches 72.8 percent, close to Claude Opus 5.5 at 73.0 percent.

Moonshot says the model is strongest in parametric modeling for mechanical parts, architectural models and configurable products. It can write and run scripts for Blender and CAD software without MCP connectors, according to the company. Moonshot has also introduced a “3D Make” feature on Kimi.com for previewing, editing and exporting 3D models.

For creators, that points to a practical niche: 3D prototyping and scripted model generation. For developers, the coding scores may make Kimi worth testing alongside coding-first tools, especially in agent workflows. The Decoder has also recently covered lower-cost coding-agent competition, including DeepSeek V4.1 Flash 1017.

The model size has not changed from K3. Moonshot lists 2.8 trillion total parameters and 104 billion active parameters, with a one-million-token context window and support for text and images.

Moonshot says token prices are down 33 percent. Kimi K3.1 Pro input costs $2 per million tokens, output costs $10 per million tokens and cache hits cost $0.20 per million tokens. The company is also offering Kimi-K3.1-Pro-Highspeed, which it says runs two to three times faster, with higher listed input and output prices.

Both Kimi K3.1 Pro and Kimi-K3.1-Pro-Highspeed are available through Kimi.ai, Kimi Work, Kimi Code and the Kimi API, according to the report.

Cybersecurity limits may shape adoption

Moonshot also discusses cybersecurity controls in its technical materials. The company says cyber-related content appeared in the broader training corpus and in reinforcement-learning task mixes, and that vulnerability research capabilities increased as post-training scaled.

For the planned open-weights release, Moonshot says it removed cyber-related data from both training phases. On its internal xdev-bench-hard benchmark, the uncleaned version solved 90.5 percent of tasks, while the cleaned Kimi K3.1 Pro solved 81.0 percent. DeepSeek-V4.1-Flash reached 85.7 percent on the same test, and Kimi K3 scored 33.3 percent.

Moonshot’s stated position is that no open-weights model should push the frontier of offensive cyber capabilities, while still improving defensive uses such as auditing older systems. For businesses, that tradeoff matters: stronger open models may be useful for coding and operations, but security policy will remain part of the buying decision.