Kimi K3.1 Pro is the newest model from Moonshot’s Kimi team, and the company’s launch post leads with a 65.7% score on Terminal-Bench 4.0. That’s ahead of GPT-6 Astra’s 58.2% and just behind Claude Opus 5.5’s 66.4%. Those are the company’s own figures, published in its release post. Here’s what the Kimi K3.1 Pro numbers say, and where the post itself tells readers to be careful.
What Kimi K3.1 Pro is, and when the weights arrive
The release calls Kimi K3.1 Pro the company’s most capable open-weight model so far, with a 1-million-token context window. A faster “Highspeed” variant runs at two to three times the speed. Both are available now in Kimi’s apps and API. Kimi Enterprise administrators have to switch the model on for their organization, because it’s off by default.
The weights are what developers are waiting for. The post says the full model weights will be released by Nov. 4. Until then, nobody outside the company can run the model or test the claims independently.
Where Kimi K3.1 Pro leads, and where it trails
On Terminal-Bench 4.0, the table in the post puts Kimi K3.1 Pro ahead of GPT-6 Astra but about 0.7 points behind Claude Opus 5.5. On BenchCAD with tools it scores 92.7%, against 96.2% for Opus 5.5, 95.9% for GPT-6 Astra and 92.6% for Fable 5.1. Without tools it scores 72.8%, against 73.0% for Opus 5.5.
So is it better? No. The fair reading is “competitive with the best closed models on these tests.” The post doesn’t claim a clean sweep, and the numbers don’t support one.
The caveats the company lists itself
Several headline results come from internal benchmarks: the Kimi Code Bench, the internal finance suite, the hallucination test and a cybersecurity test. The post offers no independent verification for them. It also says results use maximum reasoning effort, that test harnesses differ between models, and that testing proprietary models “may involve some fallback behavior.”
The hallucination figure shows why context matters. Kimi K3.1 Pro scores 8.3% on its hard hallucination test with tools, against 23.4% for its predecessor and 15.0% for Opus 5.5. GPT-6 Astra scores 5.5%. The post stresses that this is a stress test and that its rates “do not represent overall error rates in daily use.” Without tools, Kimi K3.1 Pro scores 5.4%, behind GPT-6 Astra’s 4.2%.
The cybersecurity limit is deliberate
The post says removing cyber-related training data lowered Kimi K3.1 Pro’s score on its internal cyber test from 90.5% to 81.0%, and that the model’s cyber capability “remains behind” frontier proprietary models. The company says it doesn’t aim to extend the absolute frontier of cyber capability in open weights. That’s a notable choice for an open-weight release. For comparison, see how Meta’s open weights compare on license and benchmarks, and, for another fast-moving Chinese AI company, DeepSeek’s funding and the DHS order.
Editor’s take
Treat the Kimi K3.1 Pro launch as a claim to verify, not a result. The useful moment is Nov. 4. Once the weights are public, outside researchers can rerun the public benchmarks and the hallucination test on their own hardware. Teams choosing a model for coding agents should wait for those numbers and run their own tasks. A few points on a vendor-selected benchmark rarely predict how a model behaves on a specific codebase.
