Work that requires finding the context

Agents work inside domain-specific environments rather than receiving every relevant fact in a single prompt. The task mix spans research, chip and hardware design, cybersecurity, biology, finance, law, trading, materials, alignment and software engineering.

What the percentage means

The headline averages intended task credit, balancing domains and then partners. Some tasks award partial credit; a 12% score does not mean 12% of tasks were fully solved. Partial mode is a separate view of continuous grading where available.

Compare the whole system

Harness, reasoning budget, task coverage and permitted tools shape these results. The eight-domain subset changes the task mix. Neither a leaderboard position nor this subset isolates the cause of a model’s failures.

Versions and uncertainty

This is the launch v1.0 snapshot. The source provides aggregate scores and coverage, without per-model confidence intervals in the imported data. Changes to tasks, grading or model settings need a new comparison contract.

Where this fits in an evaluation suite

A transcript that says “done” is weaker evidence than the verified final state. Track policy violations, recoveries, and invalid attempts alongside success.

Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.

A reproducible comparison

206 tasks from 14 partners across 11 equally weighted domains. Partners within a domain receive equal weight. Three attempts per task; native harnesses when available, OpenCode otherwise. Default task limit: eight hours. Intended and partial-credit scores remain distinct, with per-model coverage and missing prices preserved.

Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.

PRIMARY EVIDENCE & FURTHER READING

  1. GDPx v1.0: The General Agentic Benchmark — SemiAnalysis ↗

    SemiAnalysis · 2026

  2. GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains ↗

    Max Kan, Andrew Megalaa, Dylan Patel · 2026

Reviewed 2026-11-02. Explanation and interpretation by RLTasks. Benchmark ownership remains with its original authors.