A broader composite test
Artificial Analysis has released version 5 of its Intelligence Index, restructuring its comparison of leading artificial-intelligence models around nine evaluations. The research group says 60% of the new composite is based on private test sets or held-out answers, an effort intended to reduce the risk that models have already encountered benchmark material during training.
The index divides its total score equally among four categories: knowledge work, coding, science and general capability. Each contributes 25%. Two evaluations are new, while three existing components have been upgraded. The organization also launched a separate Progress Index designed to track capability over time on a continuous scale.
A new coding evaluation, FrontierCode 2.0, uses 132 tasks drawn from real open-source repositories. A proposed change receives credit only if it passes blocking checks and is judged suitable for a maintainer to merge. The tasks remain private, and each is attempted three times in the group's agent harness.
Terminal-Bench-Science adds 70 terminal-based scientific research tasks. Other components include HLE-Diamond, a refined set of 1,000 questions split between knowledge and reasoning; an updated AA-Omniscience dataset; and a private hard subset of the CritPt physics benchmark. Existing knowledge-work and terminal evaluations remain in the composite, while three earlier tests were removed.
Leaders vary by task and cost
Artificial Analysis reports Claude Opus 5.5 at the top of the overall index with a score of 58, followed by GPT-6 Astra at 57 and GPT-6.1 Sol at 55, all at their stated maximum effort settings. Gemini 4 Argon follows at 50. The publisher's results show that no model leads all nine evaluations, underscoring the limitations of treating a single composite number as a complete description of capability.
The reported gap between the leading model and open-weight systems widened. Kimi K3.1 Pro, whose weights were not yet released at the time of the analysis, scored 46. Among systems with released weights, DeepSeek V4.1 Flash 1017 and GLM-5.3 scored 32.
Price and reasoning effort also produced substantial differences. Among the ten highest-scoring models in the report, per-task costs ranged from $0.96 to $13.04. GPT-6.1 Sol was reported within four points of the leader while costing about one-eighth as much per index task. For Claude Opus 5.5, moving from low to maximum effort raised the score from 35 to 58 and increased the per-task cost roughly tenfold.
These figures are the benchmark publisher's measurements, not universal rankings. Results depend on task selection, weighting, harnesses, effort settings and grading. Version 5 nevertheless offers a more contamination-resistant view of agentic coding, scientific workflows and knowledge work than a composite dominated by public question sets.
