Benchmark for terminal-based agent tasks, with v4.1 updating verifiers against reward hacking.
Artificial Analysis released Coding Agent Index v2.1, upgrading FrontierCode to 2.0 and Terminal-Bench to 4.1 with stronger verifiers to resist reward hacking. Claude Fable 5.2 and Claude Opus 5.5 tie for the lead at 58, while cost per task among configurations scoring 48+ varies roughly 30x, from $1.51 to $44.22.
Reason Machines released Reason, a coding harness that keeps the prompt prefix byte-identical across turns to trigger OpenAI's prefix-caching discount, cutting repeated-token billing from $10 to $1 per million. It claims better Terminal-Bench and DeepSWE results than leading harnesses at 22% lower cost per task than OpenAI Codex.