Moonshot just shipped Kimi K3.1 Pro, the first model in its new K3.1 series and its most capable 3T-class open-weight model yet, with a million-token context window. It is available today on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, a Highspeed variant runs 2–3× faster, and the full weights go public by November 4. If you only read one sentence today: an open-weight model is now beating GPT-6 Astra on the benchmark that measures real agentic coding.
The numbers that matter
On Terminal-Bench 4.0, which runs coding agents through realistic, resource-constrained terminal tasks with strict task-specific tests, K3.1 Pro scores 65.7% — ahead of GPT-6 Astra at 58.2% and a hair behind Claude Opus 5.5 at 66.4%. On Moonshot’s own long-horizon coding bench, it scores 61.2% against K3’s 55.95%. On BenchCAD, the 3D-from-CAD benchmark, its 92.7% sits within reach of Opus 5.5 (96.2%) and Astra (95.9%), and clearly ahead of Claude Fable 5.1.
Vendor benchmarks deserve skepticism, but Terminal-Bench numbers come from tbench.ai’s public runs, and the pattern matches what the rumor mill caught early: an anonymous Arena model named Kibitz that garbled Moonshot’s tokenizer exactly like every Kimi before it.
The reward-hacking story is the real news
Buried in the tech blog is the detail I care about most. Moonshot had K3.1 Pro build a human walking simulator on top of an open-source anatomical model, 296 skeletal components, 402 muscles. When the reward shaping stopped improving arm motion, the model didn’t keep twisting the reward function, the classic agent cheat. It ran open-loop probes of the simulator, found 36 errors in the shoulder-muscle moment-arm math, corrected them, and retrained. The simulated human went from collapsing in seconds to walking 11.2 seconds on alternating strides.
That is the difference between a model that games the test and a model that debugs the world. It’s also exactly what large-scale joint RL across hundreds of domains is supposed to buy, and it’s the strongest signal in the whole release.
3D is the flex
K3.1 Pro works across Unity, Unreal, and Godot, and the demos back it up: a mythical celestial palace in Unreal Engine 5.8 with 36 million source triangles, Nanite and Lumen rendering, a MetaHuman lead. A scroll-driven Three.js car landing page with an exploded view of nearly a thousand parts. Moonshot also shipped 3D Make on kimi.ai, preview, edit, download for 3D printing, or export as video. For an agency that builds sites all day, an open-weight model with actual spatial reasoning is not a toy benchmark. It’s a headcount question.
My take
Three months ago Kimi K3 took the top open-weight coding spot and charged Sonnet-class prices for it. K3.1 Pro is the second act: same 3T-class body, massively more post-training, and a November 4 date circled for weights. DeepSeek, Qwen, and Moonshot are no longer chasing the frontier, they’re leaning on it, and the gap that used to be measured in years is now measured in weeks.
Try it free on kimi.ai today, or wait for the weights and run it yourself. Either way, watch the Terminal-Bench table. It’s getting crowded up there.