China’s Kimi K3.1 Pro comes within a point of Claude Opus 5.5 at coding, and its weights will be free

Moonshot AI launched Kimi K3.1 Pro on October 23, a post-trained version of Kimi K3 with a 1 million-token context window that the company says scores 65.7% on Terminal-Bench 4.0, within a point of Claude Opus 5.5's 66.4% and ahead of OpenAI's GPT-6 Astra at 58.2%, with full model weights promised by November 4. Moonshot priced kimi-k3.1-pro at $2 per million input tokens and $10 per million output tokens, down from $3 and $15 for K3 and half Claude Opus 5.5's $4/$20, and said it stripped cyber-related data from mid-training and reinforcement learning for the open-weight release, cutting its in-house exploit benchmark score from 90.5% to 81.0%. All benchmark figures are Moonshot's own and have not been independently verified, and the company has not said which licence the weights will carry.

Moonshot AI’s launch artwork for Kimi K3.1. Image: Moonshot AI China’s Moonshot AI says its new Kimi K3.1 Pro model is now within a point of Anthropic’s Claude Opus 5.5 on one of the toughest coding-agent tests, and it plans to hand out the full weights for anyone to download. The company launched Kimi K3.1 Pro https://www.kimi.ai/blog/kimi-k3-1-pro on October 23, calling it its “most capable 3T-class open-weight model to date”, with a 1 million-token context window. Alongside it comes Kimi K3.1 Pro Highspeed, which Moonshot says has the same abilities but runs two to three times faster. Both are live now in the Kimi app, the Kimi Work desktop app, the Kimi Code coding tool and the Kimi API. Moonshot says the full model weights will be released by November 4. How close it gets on the benchmarks K3.1 Pro is not a new base model. Moonshot describes it as a post-trained version of Kimi K3 https://madrobot.blog/2026/09/26/free-deepseek-kimi-glm-api-nvidia/ , put through far more reinforcement learning across what it calls “hundreds of downstream domains”. All the scores below are Moonshot’s own and have not been independently checked. On Terminal-Bench 4.0, which tests coding agents on real terminal tasks, Moonshot says K3.1 Pro scores 65.7% , against 66.4% for Claude Opus 5.5 and 58.2% for OpenAI’s GPT-6 Astra. It ran its own model in the Claude Code harness and took rivals’ scores from their published reports or the official leaderboard. The gap is wider on BenchCAD, which asks models to rebuild 3D objects as CAD code. With tools, Moonshot reports 92.7% for K3.1 Pro, behind Opus 5.5 96.2% and GPT-6 Astra 95.9% but just ahead of Claude Fable 5.1 92.6% . Moonshot also claims a state-of-the-art 64.4 on OfficeQA Pro v2, a financial-document test, and says K3.1 Pro’s claim-level hallucination rate on its own hardest factual-accuracy test fell from 23.4% for K3 to 8.3%. Cheaper than K3, and half the price of Opus 5.5 On the Kimi API price list https://platform.kimi.ai/docs/pricing/chat , kimi-k3.1-pro costs $2 per million input tokens and $10 per million output tokens, down from $3 and $15 for K3. The Highspeed version costs $4 and $20, and cached input is $0.20 per million tokens for both. For comparison, Claude Opus 5.5 lists at $4 input and $20 output, twice the standard K3.1 Pro price. K3.1 Pro only works in thinking mode: according to Kimi’s model documentation https://www.kimi.com/code/docs/en/kimi-code/models.html , requests with thinking turned off are rejected. In Kimi Code, the regular model needs a Plus plan or higher and Highspeed needs Pro or higher. The Kimi app is on iOS, Android, HarmonyOS and the web at kimi.ai, and Kimi Work 3.1.0 runs on Windows and Apple silicon Macs. Enterprise administrators have to switch the model on for their organisation, and the API offers zero data retention. Moonshot’s launch post does not list any country restrictions. Weights by November 4, with the hacking skills cut back The open release comes with an unusual admission. Moonshot says it found its model’s ability to find and exploit software flaws rising steadily as training scaled, so for the open-weight version it stripped cyber-related data out of both mid-training and reinforcement learning. On its in-house exploit benchmark, a version trained without those cuts solved 90.5% of tasks, while the released K3.1 Pro solved 81.0%. Moonshot says its aim is “not to release open weights that would extend the absolute frontier of cyber capability”, and that K3.1 Pro stays behind proprietary models that have been public for at least three months. That is a different line from Mistral, which pitched its own open model as doing the security work Claude and GPT-6 Astra refuse https://madrobot.blog/2026/10/06/mistral-large-4-le-chonk-1-trillion-open-weight-cybersecurity/ . Moonshot has not said which licence the K3.1 Pro weights will carry. Kimi K3’s weights on Hugging Face https://huggingface.co/moonshotai/Kimi-K3 use Moonshot’s own MIT-style licence, free for most uses but requiring a separate deal for API providers making more than $20 million over 12 months. At around 3 trillion parameters, the full model needs data-centre hardware rather than a home computer. The launch also lands weeks after OpenAI said people linked to Moonshot tried to extract its models’ hidden reasoning https://madrobot.blog/2026/09/30/openai-moonshot-ai-kimi-distillation-campaign-hidden-reasoning/ , and after Anthropic said new safeguards cut what copycat labs get out of Claude https://madrobot.blog/2026/10/13/anthropic-distillation-defenses-claude-reasoning-99-percent/ by up to 99%. Moonshot’s post does not mention either. Why it matters If Moonshot’s numbers hold up, a Chinese model anyone can download will sit within a point of the best closed US model on coding agents, at half its API price. Together with DeepSeek’s latest cheap model https://madrobot.blog/2026/10/17/deepseek-v4-1-flash-1017-beats-claude-opus-5-benchmarks-price/ , it keeps the gap between open and closed AI measured in weeks, not years. Sources: Moonshot AI, Kimi K3.1 Pro tech blog https://www.kimi.ai/blog/kimi-k3-1-pro ; Kimi API pricing https://platform.kimi.ai/docs/pricing/chat ; Kimi Code model documentation https://www.kimi.com/code/docs/en/kimi-code/models.html .