Kimi K3.1 Pro leads Artificial Analysis' command-line agent test
The Aggregate Digest — Saturday, October 24, 2026
Kimi K3.1 Pro leads Artificial Analysis' command-line agent test
Kimi K3.1 Pro does best when agents work in a terminal or through workplace software, while tests of repairing existing code still favour Claude Sonnet 5.5. Moonshot released it on Friday, and it enters the table at #12 with an Aggregate Rating of 1934, one rating point behind Claude Sonnet 5.5 and 88 above Kimi K3.
First on Terminal-Bench 4.0 in Artificial Analysis' run. The test counts the command-line tasks an agent completes. Kimi K3.1 Pro completes 67.2%, against 63.6% for Claude Sonnet 5.5 and 59.6% for Claude Opus 5.5, both at max effort. Kimi K3 completes 12.6% at max effort. On Artificial Analysis' AutomationBench, where agents pursue workplace objectives under guardrails, Kimi K3.1 Pro is second behind Gemini 4 Argon.
Artificial Analysis, October 24, 2026. Each model's best setting on this board: max with default fallback for Claude Sonnet 5.5 and Claude Opus 5.5, max for Kimi K3, xhigh for GPT-6 Astra, high for Gemini 4 Argon; none listed for Kimi K3.1 Pro.
Behind Claude Sonnet 5.5 at repairing code. LiveBench's coding tasks are bug fixes taken from real merged pull requests. Kimi K3.1 Pro scores 81.8 there, against 94.7 for Claude Sonnet 5.5. Bug Hunt Bench counts how many of 105 planted defects a model repairs in two production repositories, and at max effort Kimi K3.1 Pro scores 36 against Claude Sonnet 5.5's 51.3.
Artificial Analysis and LiveBench, October 24, 2026. Artificial Analysis rows use Claude Sonnet 5.5 at max with default fallback and list no setting for Kimi K3.1 Pro; LiveBench lists none for either. Points on each board's 0–100 scale; different tasks, not a combined score.
The gain over Kimi K3 is in agent and coding tasks. On LiveBench's browser tasks Kimi K3.1 Pro scores 80.0, against 16.0 for Kimi K3, and its LiveBench coding score rises from 59.3 to 81.8. Document work is mixed. Kimi K3.1 Pro meets every grading criterion on 31.4% of Artificial Analysis' PDF tasks, against 22.0% for Kimi K3 at max effort, but scores 86.3% against 88.7% on its questions over long documents.
Artificial Analysis and LiveBench, October 24, 2026. Artificial Analysis rows use Kimi K3 at max effort and list no setting for Kimi K3.1 Pro; LiveBench lists none for either. Points on each board's 0–100 scale; Terminal-Bench 4.0 is in the first figure.
About the same cost per task. Artificial Analysis measured $2.11 per task across its Intelligence Index for Kimi K3.1 Pro, and $2.00 for Kimi K3 at max effort.
Kimi K3.1 Pro is worth testing beside Claude Sonnet 5.5 for terminal and workflow agents. Claude Sonnet 5.5 keeps the stronger record for repairing bugs in an existing codebase.
