Artificial Analysis rebases its Intelligence Index to v4.3

Artificial Analysis announced Intelligence Index v4.3 on September 7, 2026. It upgrades Terminal-Bench from v2.1 to v4.0 with harder tasks, and replaces the tau-cubed Banking evaluation with AutomationBench-AA, its implementation of Zapier's workflow automation benchmark with a private test set. Evaluations with private questions or answers now carry 45% of the weighting, up from 40% in v4.2. Artificial Analysis calls it a subset of the changes planned for Index v5.

Why it matters

Scores dropped across the board without any model getting worse, because the yardstick changed. Any number quoted from before September 7 is on a different scale, which makes side-by-side comparisons of screenshots, vendor decks and older blog posts quietly wrong.

Who should care

Anyone choosing models on benchmark scores, or citing them in a deck

What you can do

Re-read the current index rather than trusting a saved number, and check which version any score you are quoting was measured on.