Kimi K3.1 Pro, per Artificial Analysis: a proprietary reasoning model released October 2026, #4 of 215 on the Intelligence Index with a score of 55 against a median of 27, sitting one point behind Claude Opus 5.5. Median input and output pricing ($2.00/$10.00 per 1M tokens), 1M-token context window, text and image in, text out, 90% cache discount. $2.11 per Intelligence Index task. The catch is throughput. It decodes at 77 tokens per second, slower than the 81 median and ranked #106 of 186, and it earned 4 out of 4 on Verbosity for a reason: 150M output tokens on the Index against an 87M median. It's paying for its score by thinking out loud, and you pay for the thinking at $10 per million. If your workload tolerates latency, the Pareto math is genuinely favorable; if it's user-facing, GLM-5.3 and Gemini 4 Argon land within a point at 82-85 tokens per second and comparable per-task cost. Notable that the cheap seats are closing: DeepSeek V4.1 Flash does 46 points at $0.26 per task. Two years of headroom at 2.3% the cost, as it were. The open question the page can't answer: does verbosity correlate with Intelligence Index score across the board, or did Kimi just find the benchmark's rubric and serenade it?