Fourth of 215, on somebody else’s numbers
Artificial Analysis has placed Moonshot AI’s new Kimi K3.1 Pro fourth out of 215 models on its Intelligence Index, with a score of 55. The index puts Claude Opus 5.5 first on 58, and Kimi ahead of Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon, which all score 53, and Meta’s Muse Wonder on 52. The median for comparable models is 27.
That matters because it is not Moonshot’s own scoreboard. The company’s launch post reports its own figures — 65.7% on Terminal-Bench 4.0 against GPT-6 Astra’s 58.2% and Claude Opus 5.5’s 66.4%, and 92.7% on BenchCAD with tools against 96.2% for Opus 5.5 — all of them vendor-run. The Artificial Analysis index is a third-party evaluation suite, version 4.3.2, built from ten tests including Terminal-Bench 4.0, Humanity’s Last Exam, SciCode and GDPval-AA.
Cheap to run, slow to answer, and very talkative
The index is less flattering on everything that is not intelligence. Kimi K3.1 Pro produces 73.5 output tokens per second, which ranks it 113th of 186 models and puts it below the 82-token average. Claude Opus 5.5 runs at 96 and DeepSeek’s V4.1 Flash at 195.
It is also verbose. Artificial Analysis says the model generated 150 million output tokens working through the index, against a median of 87 million — enough to rank it 92nd of 116 on verbosity. Pricing is $2.00 per million input tokens and $10.00 per million output, with a 90% cache discount, which works out at $2.11 per Intelligence Index task. Opus 5.5 costs $5.98 for the same work and Claude Fable 5.1 costs $7.63, so Kimi delivers a score three points off the leader for roughly a third of the price.
The weights are not out yet
Moonshot calls K3.1 Pro its “most capable 3T-class open-weight model to date”, with a one-million-token context window. Artificial Analysis currently lists it as a proprietary model, because the weights are not available: Moonshot says they “will be released by Nov 4”.
Until then the model exists only as a service, running on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, alongside a faster variant Moonshot calls K3.1 Pro Highspeed and says runs two to three times quicker.
What to watch
The date is the thing. If Moonshot ships the weights on schedule, a model scoring 55 on an independent index becomes downloadable, which would make it the strongest open-weight entry on that index by some distance. If the date slips, the gap between what the company calls open and what anyone can actually run stays open.
Worth noting separately: Moonshot has said it removed cyber-related training data from K3.1 Pro to reduce misuse risk. Neither the Artificial Analysis index nor the company’s published benchmarks measure what that subtraction cost.
