Artificial Analysis has released the Intelligence Index v5, a composite measure designed to evaluate language model capabilities across four primary categories: Knowledge Work, Coding, Science, and General intelligence. This version represents a significant shift in methodology, with 60% of the Index now based on evaluations using private test sets or held-out answers to mitigate data contamination. The update introduces two new benchmarks: FrontierCode 2.0, which evaluates real-world coding tasks based on whether a repository maintainer would merge the agent's changes, and Terminal-Bench-Science, which tests 70 agentic scientific research tasks within a terminal environment.
This image relates to the FrontierCode 2.0 benchmark mentioned in the first paragraph, which evaluates real-world coding tasks.
The Index weights each category equally at 25%. In the current rankings, Claude Opus 5.5 (max) leads the Index with a score of 58, followed closely by GPT-6 Astra (57) and GPT-6.1 Sol (55). The report highlights a widening gap between frontier models and open weights models; while Kimi K3.1 Pro (weights not yet released) leads the open weights category at 46, the gap between it and the leader has expanded to 12 points. Furthermore, the data reveals significant variance in cost efficiency, with the cost per Intelligence Index task ranging from $0.96 for GPT-6.1 Terra (max) to $13.04 for Claude Fable 5.1 (max).
In addition to the Intelligence Index, Artificial Analysis launched the Progress Index v1. Unlike the Intelligence Index, which is updated periodically to reflect current frontier capabilities, the Progress Index is designed to track model progress on a single continuous scale over time. It utilizes item response theory (IRT), a statistical framework that estimates a model's capability and a task's difficulty jointly. By anchoring the scale to stable frontier models—specifically o3-mini (high) at 220 and GPT-5.5 (xhigh) at 320—the Progress Index allows for the comparison of models released years apart by linking them through shared evaluation tasks. This methodology ensures that a model's score reflects its performance relative to the difficulty of the tasks it attempted, rather than the specific set of benchmarks used in a single test.