Kimi K3.1 Pro lands 0.7 points behind Opus 5.5 on Terminal-Bench, inside Claude Code, with weights by Nov 4
Moonshot shipped Kimi K3.1 Pro on October 23 and it went straight to number one on Hacker News with 573 points. The pitch is the most capable 3T-class open-weight model so far, 2.8T total and 104B active parameters, a 1M context window, live today in the Kimi app, Kimi Work, Kimi Code and the API, plus a Highspeed variant at two to three times the speed. Full weights are promised by November 4. Blog: https://www.kimi.ai/blog/kimi-k3-1-pro
The number people will quote is Terminal-Bench 4.0: 65.7 percent, against 66.4 for Claude Opus 5.5 and 58.2 for GPT-6 Astra. The detail people should quote is that it was run inside the Claude Code harness. An open-weight model now drops into the leading closed harness and lands within a point of the model that harness was built for. Kimi Code Bench 3.0 long-horizon went from 55.95 to 61.2, an internal business-operations eval went from 8.9 to 37, and claim-level hallucination on the hard with-tools set dropped from 23.4 to 8.3 percent. Artificial Analysis independently scored it 55 on its Intelligence Index, fourth behind Opus 5.5, Sonnet 5.5 and Muse Wonder Full, measured 67 percent on Terminal-Bench 4.0, the highest of any model it has run, and put the cost at $1.98 per index task at $2 and $10 per million tokens: https://artificialanalysis.ai/models/kimi-k3-1-pro
The section that will get argued about is cyber. Moonshot says it removed cyber data from mid-training and the RL mix for the open release. On its hardest offensive-dev benchmark, the unfiltered checkpoint scores 90.5 percent pass@3 and the release scores 81.0. The stated policy is to keep the open model below frontier proprietary models that have been public for at least three months. That is the first time an open-weight lab has shipped data removal as a capability lever with numbers attached, and it hands critics an easy point: 81 is still a lot. Enterprise access is off by default at launch, and zero data retention is available on the API.
For anyone building agents, the real story is the harness line. Benchmarks used to compare models on their own scaffolds. When an open model is graded in Claude Code and comes within 0.7 points, the scaffold is the shared platform and the model is a swappable part. The November 4 weights date is the one to hold them to.
← Back to all articles
The number people will quote is Terminal-Bench 4.0: 65.7 percent, against 66.4 for Claude Opus 5.5 and 58.2 for GPT-6 Astra. The detail people should quote is that it was run inside the Claude Code harness. An open-weight model now drops into the leading closed harness and lands within a point of the model that harness was built for. Kimi Code Bench 3.0 long-horizon went from 55.95 to 61.2, an internal business-operations eval went from 8.9 to 37, and claim-level hallucination on the hard with-tools set dropped from 23.4 to 8.3 percent. Artificial Analysis independently scored it 55 on its Intelligence Index, fourth behind Opus 5.5, Sonnet 5.5 and Muse Wonder Full, measured 67 percent on Terminal-Bench 4.0, the highest of any model it has run, and put the cost at $1.98 per index task at $2 and $10 per million tokens: https://artificialanalysis.ai/models/kimi-k3-1-pro
The section that will get argued about is cyber. Moonshot says it removed cyber data from mid-training and the RL mix for the open release. On its hardest offensive-dev benchmark, the unfiltered checkpoint scores 90.5 percent pass@3 and the release scores 81.0. The stated policy is to keep the open model below frontier proprietary models that have been public for at least three months. That is the first time an open-weight lab has shipped data removal as a capability lever with numbers attached, and it hands critics an easy point: 81 is still a lot. Enterprise access is off by default at launch, and zero data retention is available on the API.
For anyone building agents, the real story is the harness line. Benchmarks used to compare models on their own scaffolds. When an open model is graded in Claude Code and comes within 0.7 points, the scaffold is the shared platform and the model is a swappable part. The November 4 weights date is the one to hold them to.
Comments