Kimi K3.1 Pro: the open model that nearly matches Claude, with hacking data cut on purpose
Moonshot’s new model lands 0.7 points behind Claude Opus 5.5 on a coding benchmark and beats GPT-6 Astra. The twist: the company stripped cybersecurity data from training so it wouldn’t release an open model that’s too dangerous.
Chinese lab Moonshot has released Kimi K3.1 Pro, its most capable open-weight model so far. It’s already in the Kimi app, Kimi Code and the API, and full weights are due by November 4. It ships with a 1-million-token context window and a Highspeed variant that Moonshot says runs two to three times faster.
The headline result is coding. On Terminal-Bench 4.0, K3.1 Pro scored 65.7%. Claude Opus 5.5 scored 66.4% and GPT-6 Astra 58.2%. A model anyone will be able to download landed within a point of the best closed model on that test.
| Benchmark (Moonshot’s figures) | Kimi K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|
| Terminal-Bench 4.0 | 65.7% | 66.4% | 58.2% |
| Hallucination, hard, with tools (lower is better) | 8.3% | 15.0% | 5.5% |
| BenchCAD with tools | 92.7 | 96.2 | 95.9 |
On Moonshot’s own hallucination stress test, the rate fell from 23.4% on K3 to 8.3%. The company also claims the top score on OfficeQA Pro v2, a financial document benchmark. On several reasoning and science tests, though, it still trails the leaders.
The most interesting decision isn’t on the chart
Moonshot found that cybersecurity data in training was driving the model’s ability to find and exploit vulnerabilities. For the open release, it took that material out. The effect shows on its internal exploit benchmark: a checkpoint without the cut scored 90.5%, the final model 81%. The stated policy is not to release open weights that push the frontier of offensive capability, and to raise that ceiling only gradually.
That matters because an open model has no off switch. Once the weights are online, any filter a company puts on its API stops applying.
The caveats
Nearly every number here comes from Moonshot, and several benchmarks are internal. The announcement doesn’t state the license or the price. And running a model this size takes data center hardware, so "open" here means "other companies can host it" more than "you’ll run it on your PC."
Still, the message is hard to miss. On coding, the gap between the best closed model and the best open one is now a rounding error. For companies that need to control where their data lives, that’s a strong argument. I’d wait for independent testing before switching anything, but I’d be watching Kimi closely while I wait.
