Kimi K3.1 Pro launches as Moonshot’s strongest open-weight model yet—matching frontier coding scores, cutting hallucinations sharply, and promising full weights by November 4.
Kimi K3.1 Pro is Moonshot AI’s latest flagship model, released on October 23, 2026. It is the first model in the new K3.1 series and is described as their most capable 3T-class open-weight model to date.
Key Specs of Kimi K3.1 Pro
Architecture: Post-trained version of Kimi K3 (same base) Mixture-of-Experts (MoE) with ~2.8 trillion total parameters and ~104 billion active parameters per token.
Context window: 1 million tokens (1,048,576).
Inputs: Text, images, and video.
Reasoning: Always-on thinking mode with adjustable effort levels (low / high / max; max is default in many evaluations).
Key Features
• 1-million-token context window
• Near-Opus coding performance
• Strong 3D & CAD skills
• Sharply lower hallucination rate
• $2 / $10 API pricing
• Open weights by Nov 4
• Highspeed 2-3× faster variant
• Multimodal text + image + video
Variants of Kimi K3.1 Pro
Standard: kimi-k3.1-pro.
Highspeed: kimi-k3.1-pro-highspeed (claimed 2–3× faster output at the same intelligence level, roughly double the price).
Open weights: Full model weights scheduled for release by November 4, 2026.
Pricing (API) Kimi K3.1 Pro
Kimi K3.1 Pro: $2.00 per million input tokens / $10.00 per million output tokens (cached input ~$0.20).
Highspeed: $4.00 input / $20.00 output.
This is about one-third cheaper than the previous Kimi K3 ($3 / $15). It is also significantly cheaper than Claude Opus 5.5 and especially GPT-6 Astra.
Available now on Kimi.ai, Kimi mobile apps, Kimi Work, Kimi Code, and the Kimi API (OpenAI-compatible endpoint). Enterprise access is off by default and must be enabled by admins.
Performance Highlights
Moonshot reports strong results, particularly in coding, agentic tasks, 3D/CAD, and finance. Independent evaluations (notably Artificial Analysis) largely confirm it is near the frontier:
Artificial Analysis Intelligence Index: Score of 55 → ranked 4th overall (out of 215+ models), only 3 points behind Claude Opus 5.5 (58). This is an 11-point jump from Kimi K3.
Terminal-Bench 4.0 (agentic coding in real terminal environments): Moonshot reports 65.7% (just behind Claude Opus 5.5 at 66.4%, ahead of GPT-6 Astra at 58.2%). Independent runs have measured around 67%.
SWE-Marathon: Leads at 51.7% (ahead of Opus 5.5’s 46.3%).
OfficeQA Pro v2 (complex financial document analysis): State-of-the-art 64.4.
BenchCAD (3D reconstruction / CAD code generation): 92.7% with tools (close to Opus 5.5’s 96.2% and Astra’s 95.9%); strong without tools as well (~72.8%).
Hallucination reduction: Claim-level factual error rate on hard internal stress tests dropped sharply (e.g., from ~23.4% on K3 to ~8.3%, or even lower in some reports).
Other Notable Results
Strong gains in data analytics (e.g., 42/43 on one independent exam), vision/language auto-annotation, and long-horizon agentic work. It underperforms more on certain scientific terminal tasks relative to the top closed models.
It is particularly highlighted for full-project coding (including games in Unity, Unreal, Godot), parametric 3D modeling in Blender/CAD, financial research and modeling workflows, and long-context agentic capabilities.
Safety Note on Cyber Capabilities
Moonshot deliberately removed cyber-related training data (from both mid-training and RL stages) before the open-weight release. On their internal hard exploit benchmark, this reduced performance from 90.5% (unfiltered checkpoint) to 81.0%. They state the goal is not to extend the absolute frontier of offensive cyber capability with open weights, while still allowing gradual diffusion of dual-use skills for defensive use.
Also Read: Anthropic Claude Duet AI Agent: Features, Price and Waitlist
Final Overview
Kimi K3.1 Pro is one of the strongest near-frontier models currently available, especially notable for closing the open-vs-closed gap to just a few points on independent indexes, while offering a large context window, multimodal support, lower pricing than its predecessor, and an imminent open-weight release. It is already live for API and product use; weights arrive in about 10 days.
Also Read: Claude Haiku 5.5 Released: Features, Pricing and AI Performance

