Artificial Analysis drops GPQA Diamond in Index v4.2
Artificial Analysis has released Intelligence Index v4.2, an interim update that adds harder agentic and document-reasoning tests and doubles the weighting of private test sets to make the benchmark harder to game.
Artificial Analysis announced version 4.2 of its Intelligence Index. The company described it as an interim release that brings forward parts of its planned v5, saying the frontier has moved "so quickly in the past weeks." Index v4 launched in January, eight months earlier.
The update adds two evaluations. AA-Briefcase is an in-house agentic knowledge-work benchmark with a private test set. It covers multi-week projects with many linked tasks and thousands of source files, and it combines rubric and pairwise grading. GDP.pdf, built by Surge AI, tests reasoning over 100 PDFs spanning 4,592 pages. Answers are graded against 1,275 expert-written criteria, and a task counts as passed only if every criterion is met. GPQA Diamond has been removed because it is now saturated.
Private, held-out test sets now make up 40% of the Index weighting, twice the share in v4.1. Artificial Analysis said this share will rise again in v5. The company also fixed answer-key errors in AA-LCR, re-anchored the Elo scales for GDPval-AA and AA-Briefcase, and made SciCode's grading sandboxes more robust.
Anthropic's Claude Fable 5.1 tops the new Index. OpenAI's GPT-6 Astra is second, 4 points ahead of GPT-5.6 Sol. Meta, SpaceXAI, Moonshot/Kimi, Z.AI and Google follow, ranked by lab. GPT-6 Astra leads GDP.pdf at 33.2%, ahead of GPT-5.6 Sol at 28.2% and Claude Fable 5.1 at 26.2%. Claude Fable 5.1 and Opus 5 lead AA-Briefcase. GPT-6 Astra was also the most token-efficient model near the frontier. Anthropic, OpenAI, Meta and Z.AI share the cost-per-task Pareto frontier.
Artificial Analysis said more incremental releases will come before v5 but gave no date.