Overview
Kimi K3.1 Pro is the flagship Moonshot AI released in the early hours of October 24, 2026, the first model in the K3.1 series and a post-trained upgrade of K3. Architecture and scale are unchanged: 2.8 trillion total parameters with 16 of 896 experts active per forward pass, roughly 104 billion active parameters, a 1 million token context and text plus image input. The change sits in post-training, where Moonshot says it rebuilt nearly every layer of the stack, from the reinforcement learning framework and long-context inference to sandbox infrastructure and rollout scheduling, so each unit of RL compute converts into capability reliably.
The gains show up mostly on agentic work. Artificial Analysis gives it 55 on the Intelligence Index, fourth worldwide and 11 points above K3 at 44, with only Claude Opus 5.5 at 58 and Claude Sonnet 5.5 and Meta Muse Wonder Full at 56 ahead. Terminal-Bench 4.0 comes in at 65.7 percent by Moonshot own count, above GPT-6 Astra at 58.2 percent and 0.7 points behind Claude Opus 5.5 at 66.4; Artificial Analysis records 67.0 percent, the highest on that benchmark. GDPval-AA Elo reaches 1,846 for third place, 312 points above K3.
Pricing is the easiest part to add up. The API charges 2 dollars per million input tokens, 10 dollars output and 0.20 dollars on cache hits, roughly a third below K3 at 3 and 15 dollars, with output at about half the Claude Opus 5.5 rate. Highspeed runs 2 to 3 times faster at double the price. Full weights are promised by November 4, which would make it the highest scoring open-weight model on the index and narrow the gap between open weights and the proprietary frontier from 12 points to 3.
Key Features
- Agentic gains from post-training: Same architecture, more reinforcement learning. Moonshot says larger scale RL pulled long-horizon coding and agent behaviour toward the proprietary frontier, with its internal Kimi Code Bench 3.0 rising from 55.95 percent on K3 to 61.2 percent and AutomationBench-AA from 58 to 73 percent.
- Four reasoning levels: The K3.1 series supports low, medium, high and max, with medium new in this release. Compute follows task complexity, so simple requests do not burn a top-tier budget and long-horizon work can step up.
- Terminal and long-horizon coding benchmarks: Terminal-Bench 4.0 lands at 65.7 percent by Moonshot count and 67.0 percent as recorded by Artificial Analysis, the highest on the benchmark and above GPT-6 Astra at 58.2. SWE-Marathon v1.1 reaches 51.7 percent for first place on multi-hour change sets.
- 3D and professional workflows: It writes and runs Blender and CAD scripts directly, generates complex Three.js scenes and edits projects in Unity, Unreal Engine and Godot. BenchCAD scores 92.7 percent with tools, 3D Make ships in the consumer app, and OfficeQA Pro v2 reaches 64.4 for financial documents.
- Pricing tiers and Highspeed: Standard pricing is 2 dollars input and 10 dollars output per million tokens with 0.20 on cache hits, about a third below K3 on all three rates. The Highspeed variant runs 2 to 3 times faster at double the price for latency-sensitive work.
- Weights and training data handling: Full weights are expected by November 4. Moonshot says it found cyber-related material in general training data and removed it during mid-training and RL while keeping the general coding tasks that make the model useful.
Use Cases
- Repository-wide refactors and long-horizon tasks, running multi-hour change sets at a high reasoning level and reviewing the diff afterwards.
- Agents that drive real terminal work, from builds to ops scripts, where Terminal-Bench style tasks are its strongest range.
- Teams that need self-hosting, deploying the weights on their own cluster once they ship on November 4 so code and data stay inside.
- 3D and game prototyping, generating Three.js scenes or editing scripts and assets inside Unity and Godot projects.
- Financial research workflows, using the document question answering and table checks behind OfficeQA Pro v2 on long filings.
- Latency-sensitive products on the Highspeed tier, trading a higher unit price for 2 to 3 times the output speed.
Pros
- A score of 55 puts it fourth worldwide, 11 points above K3 and within 3 points of the proprietary frontier.
- Terminal-Bench 4.0 at 65.7 percent puts it ahead of GPT-6 Astra on terminal agent tasks.
- All three API rates run about a third below K3, with output at roughly half the Opus 5.5 price.
- Full weights are promised by November 4, giving a clear path to self-hosting.
- Four reasoning levels let compute follow task complexity instead of defaulting to the top tier.
- The Highspeed variant offers 2 to 3 times the output speed as a separate tier.
Pricing
API pricing runs per token: 2 dollars per million input, 10 dollars output and 0.20 dollars on cache hits. All three rates sit about 33 percent below K3 at 3 and 15 dollars, and output costs roughly half the Claude Opus 5.5 rate. Highspeed costs double the standard price for 2 to 3 times the output speed. The model is available through Kimi.com, Kimi Work, Kimi Code and the Kimi API, and full weights are expected by November 4, 2026, so self-hosting carries only compute cost. Check the official site for current quotas and rate limits.
Summary
K3.1 Pro spends compute on post-training rather than parameter count. The architecture is K3, yet the intelligence index jumps from 44 to 55, Terminal-Bench 4.0 reaches 65.7 percent and prices fall by a third, and with those three holding at once an open-weight model sits 3 points from the proprietary frontier for the first time. Teams running long-horizon coding and terminal agents get the clearest cost case on the market today, and the arithmetic changes again when the weights land on November 4.
Version History
- Kimi K3.1 Pro launches as a post-trained K3, fourth on the AA Intelligence Index at 55 (2026-10-24): Moonshot released Kimi K3.1 Pro in the early hours of October 24, 2026, the first model in the K3.1 series and a post-trained version of K3 that keeps the 2.8 trillion parameter, 104 billion active, 1 million token architecture. Artificial Analysis scores it 55 on the Intelligence Index for fourth worldwide, 11 points above K3; Terminal-Bench 4.0 reaches 65.7 percent by Moonshot count and SWE-Marathon v1.1 takes first at 51.7 percent. Pricing is 2 dollars input and 10 dollars output per million tokens, about a third below K3 on all three rates, with a Highspeed tier at double the price and double the speed. Full weights are promised by November 4, and Baidu Dazi Enterprise added it to its official model list on October 25.