The index combines FrontierCode v1.1, Terminal-Bench 4.0, and SWE-Atlas-QnA-AA with equal weight. The source lists 100 FrontierCode tasks, 66 Terminal-Bench tasks, and 124 technical Q&A tasks in SWE-Atlas-QnA-AA.
Scores are the mean of three attempts per task. The published release summary identifies Claude Opus 5.5 running in Claude Code as the leading configuration, with a score of 61.7. The index page also presents per-benchmark results alongside API cost per task, token use, active agent wall time, safety-refusal rates, and reward-hacking results.
Artificial Analysis describes its cost measure as pay-per-token API cost, including cache pricing where applicable. It does not represent consumer subscription prices or full deployment cost; infrastructure, engineering, and supervision are excluded. Active runtime also excludes environment startup, verifier or judge time, and other harness overhead.
The index is a comparative measurement under its selected tasks, agent settings, and scoring rules. The source does not establish that the top-ranked configuration will perform best, cost least, or be most reliable in a particular production repository or workflow.