started · updated
Artificial Analysis updates AI Intelligence Index to version 5
Artificial Analysis has released version 5 of its Intelligence Index (AAII), a significant restructuring designed to prevent model performance inflation. To combat “data contamination” where models train on public benchmarks, 60% of the evaluation now relies on private test sets or hidden answers.
The updated index incorporates new specialized benchmarks, including FrontierCode 2.0, which evaluates coding ability based on whether code changes are practical enough for a maintainer to merge into real-world repositories, rather than just whether the code runs. It also includes Terminal-Bench-Science for scientific research tasks and HLE-Diamond for complex reasoning and knowledge.
In the overall rankings, Google’s Gemini 4 Carbon holds the top position with 59 points, followed by Anthropic’s Claude Opus 5.5 (58 points) and OpenAI’s GPT-6 Astra (57 points). While Chinese models like Kimi K3.1 Pro show strength in specific areas, the gap between top-tier closed-source models and released open-weight models has widened under this more rigorous testing framework.
The update also highlights the massive performance and cost variance between different “effort levels” within the same model, noting that the gap between low-cost and high-performance modes is expanding significantly.