Moonshot shipped Kimi K3.1 Pro yesterday, and it's the most interesting open model release of the month. The Beijing lab calls it the first of a new K3.1 series and its most capable 3T-class model to date. It's a post-trained Kimi K3: the same 2.8-trillion-parameter mixture-of-experts architecture with 104 billion active per token, the same 1-million-token context, but with far more reinforcement learning compute across what Moonshot says are hundreds of downstream domains.

The headline numbers come from Moonshot's own testing, so read them with that in mind. On Terminal-Bench 4.0, K3.1 Pro scores 65.7%, ahead of GPT-6 Astra's 58.2% and just 0.7 points behind Claude Opus 5.5 at 66.4%. It also takes Moonshot's own long-horizon coding test, Kimi Code Bench 3.0, at 61.2% versus 55.95% for K3. On BenchCAD, a 3D-reconstruction test, it hits 92.7% with tools, right on the heels of Opus 5.5's 96.2%.

Priced like a mid-tier model, benchmarked like a flagship

Here's the part that should make competitors nervous. The API costs about $2 per million input tokens and $10 per million output tokens, with cached input at $0.20. That's roughly a third cheaper than Kimi K3's $3 and $15. A Highspeed variant runs two to three times faster for $4 and $20. If you'd rather not touch an API, it's already selectable in the Kimi app on web, iOS, Android and HarmonyOS.

The independent read backs up the hype. Artificial Analysis ranks K3.1 Pro fourth on its Intelligence Index with a score of 55, behind only Claude Opus 5.5, Claude Sonnet 5.5 and Meta's Muse Wonder Full. It notes the cost per completed task sits near $2, in line with what K3 charged, because the model writes more output tokens even though each token costs less.

Open weights are the real story

The weights aren't public yet. Moonshot promises the full release by November 4, 2026. If it delivers, Artificial Analysis says K3.1 Pro will be the highest-scoring open-weights model on its index, narrowing the open-versus-closed gap from 12 points to 3. A 2.8-trillion-parameter model won't run on consumer hardware, so expect the usual cloud and community hosting options to appear quickly.

Our take: this is the best value proposition in open AI right now, provided the weights actually ship. Two honest caveats. Moonshot deliberately removed cyberattack-related training data before this release, so don't expect it to be a security-research tool. And almost all the benchmark figures are Moonshot's own. We'll know much more once independent evaluators can run the weights themselves, sometime after November 4.