Moonshot’s Kimi K3.1 Pro lands 3 points behind Claude Opus 5.5

On the Artificial Analysis Intelligence Index the model scores 55, up 11 points on K3 and fourth overall. Three models beat it: Claude Opus 5.5 (58), and Claude Sonnet 5.5 and Meta’s Muse Wonder Full (56 each). If the weights arrive on time, the best open model would sit 3 points off the top, down from 12, The Decoder calculates.

The biggest gains are in agent work. On Terminal-Bench 4.0, which scores coding agents working at a command line, Moonshot reports 65.7 per cent, The Decoder notes. Opus 5.5 scores 66.4 and OpenAI’s GPT-6 Astra 58.2. Developers pay a third less through the API for a model of the same size, The Decoder adds: 2.8 trillion parameters, the internal settings that store what it learned. A Highspeed version runs two to three times faster, Moonshot says.

Claude Pro, Anthropic’s paid plan for people, costs $20 a month, or $17 a month billed yearly, on our price sheet, read 23 October 2026.

Once the weights ship, the nearest open rival to Opus 5.5 on this index trails it by 3 points and costs nothing to download, while Claude Pro costs $20 a month, so the bill for an open model moves to whoever pays to run 2.8 trillion parameters.