Kimi Releases K3.1 Pro, Beats GPT-6, Returns as Top Open-Source Model

Moonshot AI released Kimi K3.1 Pro on October 23, ending more than three months without a major flagship update from China’s leading large-model developers. The launch follows a packed September in which OpenAI introduced GPT-6 Astra and GPT-6 Sol, while Anthropic unveiled Claude Opus 5.5 and Claude Sonnet 5.5. K3.1 Pro is a 2.8 trillion-parameter model that activates 104 billion parameters per inference and supports a 1 million-token context window. It targets programming, agents, 3D creation, and financial research. Moonshot said the full model weights will be available by November 4. In official benchmarks, K3.1 Pro posted 65.7 percent on Terminal-Bench 4.0, ahead of GPT-6 Astra’s 58.2 percent and slightly behind Claude Opus 5.5’s 66.4 percent. It scored 51.7 percent on SWE-Marathon v1.1, surpassing Claude Opus 5.5 and GPT-6 Astra. On Agents’ Last Exam, it reached 43.2 points, and it topped the tool-enabled version of Humanity’s Last Exam with 70 percent. GPT-6 Astra and Claude Opus 5.5 still lead in some scientific-computing, reasoning, and 3D-reconstruction tasks. Independent evaluator Artificial Analysis gave K3.1 Pro a score of 55 on its latest intelligence index, placing it fourth globally behind Claude Opus 5.5, Claude Sonnet 5.5, and Muse Wonder Full, and 11 points above the previous K3. Once its weights are public, Artificial Analysis expects K3.1 Pro to become the highest-scoring open-weight model, narrowing the gap between leading open and closed systems from 12 points to 3. Moonshot emphasized stronger performance on complex, long-running workflows. Demonstrations included coding Unreal Engine 5.8 and Godot-based games, generating brand-style 3D web pages, and controlling Blender and CAD workflows to build editable assets. A new 3D Make feature allows users to create, preview, edit, export, or print models from Kimi’s web and mobile products. In a prolonged human gait simulation, the model worked through 167 tasks and 420 reviews, eventually fixing calculation errors and enabling stable walking. The gains come from extensive post-training rather than training a new 3-trillion-parameter model from scratch. Moonshot increased reinforcement-learning compute and adjusted its long-context reasoning, sandboxing, and task-scheduling systems. Financial research also improved, with K3.1 Pro scoring 64.4 on OfficeQA Pro v2, ahead of Claude Opus 5.5 and GPT-6 Astra. Internal tests showed lower factual-error rates under stress, falling from 23.4 percent to 8.3 percent. K3.1 Pro is priced at $2 per million input tokens and $10 per million output tokens, about one-third lower than K3. A Highspeed version promises two to three times faster output at double the cost, though Artificial Analysis measured standard output at 73.5 tokens per second, below some peers. The model is now available across Kimi’s products and API.