GDPVal-AA v2

GDPVal-AA v2

Quick answer: GDPVal-AA v2 is Artificial Analysis's benchmark for measuring AI model performance on real-world, economically valuable tasks across a wide range of occupations. Scores are Elo ratings anchored to a human baseline of 1,000 — higher scores indicate performance above the human baseline. It is one of the nine evaluations in the Artificial Analysis Intelligence Index v4.1. Claude Fable 5 leads at 1,760 Elo.

At a Glance

What it tests: An AI model's ability to complete real-world work tasks that have direct economic value — the kind of tasks performed by knowledge workers, professionals, and specialists across occupations. Tasks are drawn from realistic work scenarios rather than narrow academic exercises.

Why it matters: Most intelligence benchmarks test reasoning or knowledge in isolation. GDPVal-AA v2 specifically targets economically productive output — aligning AI evaluation with real-world utility. By anchoring to a human baseline of 1,000, the scores give intuitive meaning: a score of 1,300 means substantially better than a typical human at these tasks, while 900 means somewhat below.

Known limitations: GDPVal-AA v2 is a proprietary benchmark developed and maintained by Artificial Analysis. Methodology details are published at artificialanalysis.ai/methodology/intelligence-benchmarking. The benchmark tasks and exact scoring methodology are not open-source.

What GDPVal-AA v2 Measures

GDPVal-AA v2 evaluates AI models on real-world, economically valuable tasks across a wide range of occupations. This includes tasks like:

• Writing, summarizing, and editing professional documents
• Analyzing data and producing actionable recommendations
• Planning, scheduling, and workflow management
• Research and information synthesis
• Professional communication and correspondence

Each task is scored on whether the model's output would be considered useful and correct by a domain expert, generating Elo ratings through comparative evaluation. The score is anchored so that the human baseline = 1,000.

The Intelligence Index uses a normalized form of the score: (Elo - 500) / 2000, but the raw Elo values reported by model cards are the standard comparison metric.

Benchmark Specifications

How GDPVal-AA v2 Is Scored

Models are evaluated on a set of real-world work tasks. Outputs are compared against a human baseline and against each other using Elo-style comparison, producing a rating anchored so that an average human worker scores approximately 1,000. Scores above 1,000 indicate superhuman performance on the task distribution; scores below 1,000 indicate below-human performance.

State-of-the-Art Results

Scores from Inkling model card (Thinking Machines Lab, July 2026). All 9 comparison models reported.

Human baseline = 1,000.

GDPVal-AA v2 vs Related Benchmarks

Last updated 2026-07-16.