Source: Artificial Analysis research publication by George Cameron and Micah Hill-Smith, published October 28, 2026, titled 'Announcing Artificial Analysis Intelligence Index v5'.
The comprehensive overhaul was prompted by widespread training set leakage and benchmark saturation across public academic suites. In the revamped coding quadrant, Cognition's newly minted FrontierCode 2.0 presents 132 real-world repository challenges verified against live containerized execution environments. Claude Opus 5.5 led FrontierCode 2.0 at 46% pass rate, while GPT-6 Astra demonstrated superiority on Terminal-Bench-Science with 63% completion and on the private 40-question CritPt v2 physics subset with 72% accuracy.
Task token consumption expanded drastically under v5.0, with agentic environments demanding long-horizon planning and recursive execution traces. Terminal-Bench-Science alone accounted for approximately one-third of all generated output tokens across an evaluation pass. Reasoning effort settings demonstrated profound impacts on output quality: Claude Opus 5.5 improved its composite score from 35 at low reasoning effort to 58 at maximum effort, though accompanied by a ten-fold increase in cost per completed benchmark task.
The newly launched Progress Index v1 standardizes progress measurements back to November 2022 by normalizing evaluation across multi-year point-in-time snapshots. Developed with methodological guidance from Epoch AI and Anthropic's research divisions, the index decouples benchmark difficulty escalation from underlying algorithmic capability gains, providing enterprise architects with an absolute historical velocity trajectory.