Skip to content
Moojaz — موجز
Search
Menu
HomeRegionsMaghrebAfricaMiddle EastEuropeAsiaAmericasWorldSectionsPoliticsEconomySecurityHealthScienceTechnologyAIEnvironmentCultureSportsSearch
File

GDPx: evaluating AI agents in specialized work

This dossier follows SemiAnalysis’s GDPx benchmark, its methods and results across eleven specialized domains. Scores are those published by the benchmark authors.

Last updated : 1 key moments
Generic artificial neural network diagram.
Cburnett diagram (2006), CC BY-SA 3.0, PNG rendering; general illustration, not the architecture of evaluated models. — Cburnett / Wikimedia Commons — CC BY-SA 3.0 https://creativecommons.org/licenses/by-sa/3.0/ — schéma de 2006, rendu PNG.

Where things stand

As of 2 November 2026, version one comprises 206 tasks from 14 partners. The published ranking puts Claude Opus 5.5 at 12%, a score that requires methodological context.

Reported information

GDPx: new 206-task benchmark puts the top model at 12%

SemiAnalysis published GDPx on 2 November 2026, evaluating AI agents on 206 tasks across eleven domains including chip design, cybersecurity, biology, law and coding. Its initial ranking gives Claude Opus 5.5 12%, followed by Gemini 4 Carbon at 10.9% and GPT-6 Astra at 9.6%.

Generic artificial neural network diagram.
Cburnett diagram (2006), CC BY-SA 3.0, PNG rendering; general illustration, not the architecture of evaluated models. — Cburnett / Wikimedia Commons — CC BY-SA 3.0 https://creativecommons.org/licenses/by-sa/3.0/ — schéma de 2006, rendu PNG.

Tasks come from 14 specialist partners. The overall score weights domains equally and combines binary or partial grading depending on the task. It therefore does not directly represent the share of jobs or business assignments that can be automated.

The authors say safety mechanisms can reduce some scores, particularly in cybersecurity, biology and alignment. Results depend on testing conditions and the tools provided to agents.

These are SemiAnalysis’s figures, rather than independently reproduced performance measurements.

Sources 4