- Published on
Artificial Analysis Intelligence Index v4.2
- Authors
- Name
- DeepHorus
On September 4, 2026, the creators of the Artificial Analysis Intelligence Index released version 4.2, an interim update designed to keep the benchmark relevant amid the rapid advancement of frontier AI models. This update introduces more complex, realistic evaluation tasks and significantly increases the use of private test sets to prevent model evaluation gaming. The Index aims to provide a more accurate reflection of real-world AI capabilities.
The update arrives eight months after the launch of v4 in January, prompted by the accelerating pace of new model releases. The team deliberately held back changes to maintain stability through major launches but felt an immediate update was necessary to ensure the Index remained a useful tool. The v4.2 release is a precursor to the full v5 version, which is already under development.
Key additions to the evaluation suite include AA-Briefcase, an in-house test assessing models on multi-week knowledge work projects with thousands of source files, and GDP.pdf, a long-context document reasoning challenge spanning 4,592 pages of PDFs. The weighting of private, held-out test sets has been doubled to 40% of the total Index, significantly reducing opportunities for developers to tailor models to specific public benchmarks.
The updated results show a competitive frontier, with Anthropic’s Claude Fable 5.1 leading the Index, followed by OpenAI’s GPT-6 Astra. Meta ranks third, ahead of other labs like SpaceXAI and Google. Notably, GPT-6 Astra demonstrates superior token efficiency compared to other leading models near the intelligence frontier.
The revised Index also highlights a shared Cost per Task frontier among four labs: Anthropic, OpenAI, Meta, and Z.AI. Improvements to grading infrastructure, such as enhanced scoring accuracy for the AA-LCR evaluation and more stable Elo scaling, underpin the new version’s reliability. The team states the focus is on measuring genuine utility and preventing score inflation.
Looking ahead, the developers plan more incremental releases leading up to the major v5 overhaul. The increased reliance on private test sets in v4.2 is expected to grow further in the next version, signaling a continued commitment to creating more robust and anti-gaming evaluations for the fast-evolving AI industry.
