Moonshot AI introduced Kimi K3.1 Pro, a new open-weight model matching or nearing the world's strongest proprietary AI models in coding, agentic work, and knowledge tasks. In company benchmarks, K3.1 Pro achieves 65.7% on Terminal-Bench 4.0, ranking ahead of OpenAI's GPT-6 Astra at 58.2% and just behind Anthropic's Claude Opus 5.5 at 66.4%.

The Chinese AI developer stated that K3.1 Pro is the inaugural model in the K3.1 line and its most capable 3T-class open-weight model to date. Featuring a 1-million-token context window, it delivers major improvements in coding, 3D creation, financial research, and agent operations, while offering lower API pricing than its predecessor.

Moonshot also launched Kimi K3.1 Pro Highspeed, running two to three times faster than the standard version. Both models are available now on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, with full model weights scheduled for release by November 4, 2026.

Same Giant Brain, Far More Training

K3.1 Pro uses the architecture of Kimi K3, the 2.8-trillion-parameter model launched by Moonshot in July, but incorporates significantly more post-training compute.

Moonshot executed large-scale reinforcement learning across hundreds of domains simultaneously by optimizing its reinforcement learning framework, long-context inference, training efficiency, sandbox infrastructure, and rollout scheduling. This generated consistent performance gains across every tested domain.

Where It Beats Claude and GPT

Moonshot evaluated K3.1 Pro against Claude Opus 5.5, Claude Fable 5.1, GPT-6 Astra, DeepSeek-V4.1-Flash-1017, and GLM-5.3 using self-reported results with K3.1 Pro at maximum reasoning effort.

K3.1 Pro led several evaluations, including SWE-Marathon with 51.7 compared to Claude Opus 5.5 at 46.3, and Humanity's Last Exam Full with tools at 70.0 versus 67.7 for Opus 5.5. It also led OfficeQA Pro v2 at 64.4, Agents' Last Exam at 43.2, GDP.pdf at 34.6, and ZeroBench with tools at 63.0. Moonshot reported leads on BrowseComp at 92.1, AutomationBench-AA at 73.3, and DeepSWE v1.1 among proprietary models.

It also surpassed GPT-6 Astra on the GDPval-AA v2.1 work-task ranking with an Elo score of 1,846, remaining slightly behind Claude Opus 5.5 at 1,866.

Where It Still Falls Behind

The model trails in specific categories, scoring 38.1 on Terminal-Bench Science 0.1, behind GPT-6 Astra at 68.1 and Claude Opus 5.5 at 63.3.

It also trails top proprietary systems on ProgramBench with 31.5 against 65.0 for Opus 5.5, HLE-Diamond with 39.4 against 65.8 for GPT-6 Astra, and GPQA-Diamond with 94.2 against 96.1 for GPT-6 Astra. On cybersecurity benchmarks like ExploitBench and ExploitGym, it stays behind Claude and GPT, which Moonshot described as intentional.

Against other open models, K3.1 Pro maintains a clear lead, scoring 65.7 on Terminal-Bench 4.0 while DeepSeek-V4.1-Flash-1017 scores 42.3 and GLM-5.3 scores 41.8.

Builds Games and Fixes Own Mistakes

Moonshot reported that K3.1 Pro handles tasks from quick code fixes to full project builds, integrating with game engines like Unity, Unreal Engine, and Godot alongside web search and 3D generation.

In an internal test, the model built a human walking simulator using an open-source anatomical model containing 296 skeletal and 402 muscle components. Across 167 missions, it progressed the simulated human from collapsing instantly to walking continuously for 11.2 seconds, identifying and correcting 36 errors in shoulder-muscle calculations before retraining.

It Now Builds 3D Models Too

K3.1 Pro scores 92.7% on BenchCAD with tools for rebuilding 3D objects from images using CAD code, approaching Claude Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%. Without tools, it scores 72.8, nearly matching Opus 5.5 at 73.0.

The system converts text briefs, sketches, or reference images into usable 3D models and runs scripts for Blender and CAD software without MCP connectors. Moonshot introduced a 3D Make feature on Kimi.ai and the Kimi app for previewing, editing, and downloading 3D models for printing.

Big Jump in Finance and Accuracy

Moonshot stated K3.1 Pro executes end-to-end financial workflows, scoring 58.6 on Frontier Finance to lead Opus 5.5 at 56.7 and GPT-6 Astra at 45.0.

Regarding factual accuracy, the claim-level hallucination rate on the internal Kimi Hallu Bench Hard test dropped from 23.4% for K3 to 8.3% for K3.1 Pro, outperforming Claude Opus 5.5 at 15.0% but trailing GPT-6 Astra at 5.5%.

How It Compares With Kimi K3

On long-horizon coding, K3.1 Pro scores 61.2% on Kimi Code Bench 3.0, up from 55.95% for K3. Internal finance scores rose from 50.4 to 57.6 in financial modeling and from 58.8 to 70.4 in financial research, while simulated business operations rose from 8.9 to 37.0.

Both systems use a 1-million-token context window, with pricing serving as the primary practical difference for developers.

A Third Cheaper Than Kimi K3

Moonshot API pricing sets K3.1 Pro at $2 per million input tokens, $10 per million output tokens, and $0.20 per million cached input tokens.

This represents a one-third reduction compared to Kimi K3, which costs $3 per million input tokens, $15 per million output tokens, and $0.30 for cached input. The K3.1 Pro Highspeed version costs $4 per million input tokens and $20 per million output tokens.

Pricing remains flat across the 1-million-token context window with no extra prompt charges. API access requires a minimum $1 top-up, and Zero Data Retention applies for API customers.

Deliberately Limited Hacking Skills

Moonshot removed cyber-related training data from mid-training and reinforcement learning for the open-weight release after observing it increased vulnerability exploitation abilities.

An unrestricted version solved 90.5% of tasks on the internal xdev-bench-hard test, while the final K3.1 Pro solves 81.0%, maintaining lower performance than frontier proprietary models in offensive cybersecurity.

How to Use It

K3.1 Pro rolls out on Kimi.ai, Kimi mobile apps for iOS, Android, and HarmonyOS, Kimi Work desktop app version 3.1.0 or later, and Kimi Code. Developers can access it via the Kimi API using model IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed.

Kimi Enterprise administrators can enable the model, though it is disabled by default at launch. Full open weights release by November 4, 2026.