The interesting number in Moonshot’s latest release is not the headline score. It’s the gap. On Terminal-Bench 4.0, Kimi K3.1 Pro lands at 65.7%, less than a point behind Claude Opus 5.5 and about seven points ahead of GPT-6 Astra. An open-weight model is now sitting in the same tier as the best closed ones, and that changes the buying conversation.
Moonshot announced K3.1 Pro as the first model in its K3.1 series. It is a post-trained version of Kimi K3, a 3T-class model with a 1-million-token context window, built with much more post-training compute than its predecessor.
What changed under the hood
Moonshot says the gains come from reinforcement learning across hundreds of downstream domains at once. To make that work, the team reworked the RL framework, long-context inference, sandbox infrastructure, rollout scheduling, and the tooling that checks environment quality. The stated goal is that each added unit of RL compute produces predictable returns. That is a claim about process, not a benchmark, and we can’t verify it. But the results are consistent across domains, which is what you’d expect if it holds.
There is also a Highspeed variant, K3.1 Pro Highspeed, which Moonshot says runs 2 to 3 times faster for interactive use. Agent loops live or die on latency, so this matters more than it sounds.
Coding and 3D work
Coding is the main pitch. On Kimi Code Bench 3.0 Long-horizon, an internal benchmark, K3.1 Pro scores 61.2% against K3’s 55.95%. Internal benchmarks deserve skepticism, so weigh the Terminal-Bench result more heavily.
The demos lean on game engines and 3D. Moonshot says the model built an Unreal Engine 5.8 scene with a procedurally modeled dragon, and a scroll-driven Three.js car showcase with 968 mechanical parts. It also worked through a long simulation task, a human walking model, over 167 missions. When tuning rewards failed, it probed the simulator and fixed 36 errors in muscle calculations. That is the behavior people want from coding agents: stepping back instead of looping on the same bad fix.
On BenchCAD, K3.1 Pro scores 92.7% with tools, behind Claude Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%. It writes scripts for Blender and CAD tools directly, without MCP connectors. A new 3D Make feature on kimi.ai lets you preview, edit, and export models for 3D printing.
Finance and hallucination claims
Moonshot also pushes finance. K3.1 Pro posts 64.4 on OfficeQA Pro v2, which it calls state of the art, and improves on K3 across internal tests for modeling, research, and business simulation. The most notable figure is the claim-level hallucination rate on its Hallu Bench Hard, down from 23.4% to 8.3%. Moonshot is careful to say this is a stress test, not an everyday error rate. Read it that way.
Availability and why it matters
K3.1 Pro and Highspeed are live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. Full weights are promised by Nov 4. Pricing was not included in the announcement, which is a gap for anyone budgeting.
So does it matter? Yes, with caveats. If the weights land on schedule, teams that need self-hosting, auditability, or lower costs get a near-frontier option. Rivals like GLM-5.3 and DeepSeek-V4.1-Flash are the obvious comparisons, and Moonshot plotted both against cost per task. But a 3T-class model is not something most shops will run themselves, so API pricing will decide how much of this is real. Watch independent evaluations once the weights are out.
