Artificial Analysis has launched the fifth version of its Intelligence Index, marking the most significant update to its composite model capability measure since the release of v4 in January. The v5 update introduces two new benchmarks and three major upgrades, with 60% of the index now derived from private datasets or held-out answers to better reflect real-world performance.

New Benchmarks and Upgrades

The v5 update incorporates two new evaluation tools designed to test agentic capabilities:

  • FrontierCode 2.0: Developed by Cognition, this benchmark evaluates coding tasks within real open-source repositories. Performance is graded based on whether a repository maintainer would accept the agent's proposed changes.
  • Terminal-Bench-Science: This test consists of 70 scientific research tasks performed within a terminal environment, requiring agents to successfully navigate complex workflows.

Additionally, the index has been refined through three key upgrades. HLE-Diamond replaces the previous "Humanity's Last Exam" with a 1,000-question set balanced between reasoning and knowledge. AA-Omniscience v1.1 has been updated to remove ambiguous or saturated questions, and CritPt v2 now utilizes a private, 40-question "hard" subset that requires the use of code-execution tools.

Performance and Cost Analysis

The current leaderboard shows Anthropic and OpenAI models occupying the top five positions. Claude Opus 5.5 (max) leads the index with a score of 58, followed by GPT-6 Astra (max) at 57 and GPT-6.1 Sol (max) at 55.

The gap between proprietary models and open-weights models has widened to 12 points. While Kimi K3.1 Pro leads the open-weights category with a score of 46, released open-weights models like DeepSeek V4.1 Flash 1017 and GLM-5.3 currently trail at 32.

Cost efficiency remains a significant differentiator among top-tier models. The analysis found a 13x variance in cost per task among the leading models. For instance, GPT-6.1 Sol (max) provides performance within four points of the leader, Claude Opus 5.5 (max), while costing approximately one-eighth as much per task. Furthermore, the data indicates that "effort settings"—the amount of compute and tokens dedicated to a task—have a more pronounced impact on performance in v5 than in previous versions.

Introducing the Progress Index

Alongside the Intelligence Index, Artificial Analysis has introduced the Artificial Analysis Progress Index v1. While the Intelligence Index remains the primary tool for assessing current model capabilities, the Progress Index tracks model performance on a continuous scale dating back to November 2022. This new tool is intended to help users compare how underlying AI capabilities have evolved over time, regardless of when a specific model was released.

The v5 update highlights that no single model currently dominates all nine evaluation categories, leaving substantial room for future development. As models become more agentic, the demand for output tokens has risen, with the median token usage per task increasing by two-thirds compared to the previous version.