SemiAnalysis published its GDPx leaderboard on 2 November 2026. The benchmark contains 206 agent tasks across 11 domains, developed with 14 specialist partners. Claude Opus 5.5 leads the initial public table with a 12.0% composite score, followed by Gemini 4 Carbon at 10.9% and GPT-6 Astra at 9.6%.
That 12.0% is not a simple count of tasks completed. GDPx gives each domain equal weight, gives partners equal weight within a domain, and combines binary and continuous grading. Its displayed results average three attempts per task. A reader comparing models should start with those scoring rules, not treat the headline percentage as a universal success rate.
What the tasks test
The published methodology describes agents working in environments with tools, files and constraints that resemble specialist jobs. Domains include software engineering, chip and hardware design, cybersecurity, biology, finance, legal work and materials science. The default task time limit is eight hours, although some tasks have shorter limits. SemiAnalysis says it used each model’s native agent harness where available.
The distinction matters because a short question and a long assignment test different abilities. In one public software engineering task, for example, an agent must change Git LFS behavior in a partial clone while preserving behavior in a full clone. The benchmark evaluates the resulting work, including requirements that are easy to miss when a local fix appears to succeed.
Reading the result carefully
The initial leaderboard shows variation by domain and by cost per task. SemiAnalysis also says safety classifiers can affect results in cybersecurity, biology and alignment. Its finance and legal tasks are difficult enough that the authors report partial credit in the headline comparison where their intended binary grading produced zero scores.
GDPx is an evaluation designed and reported by SemiAnalysis with its partners. Its aggregate score does not establish how a model will perform in a particular company’s repository, laboratory or approval process. For an operator choosing an agent, the useful next step is to inspect the relevant domain, task design, grading rule, harness and cost, then test the model against the work and controls it would actually encounter.
