GDPx gives leading AI agents a harder test—and a low opening score

The best opening result on GDPx is 12%. That is a striking number for a new AI benchmark, but it needs a careful denominator. It does not mean the leading model can do 12% of all human work, or that it passed exactly 12% of a uniform collection of questions.

SemiAnalysis announced GDPx on November 2, 2026. Its version 1.0 site lists 206 tasks across 11 domains, supplied by 14 specialist partners. The overall table currently places Claude Opus 5.5 at 12.0, Gemini 4 Carbon at 10.9 and GPT-6 Astra at 9.6. These are the benchmark creator’s reported results, not independent measurements by UpThat.

A deliberately selective view of work

GDPx targets areas SemiAnalysis expects to generate high returns from frontier AI, rather than sampling occupations in proportion to their share of the economy. That is the firm’s selection thesis, not an established forecast of where AI revenue will appear.

The domain catalogue makes the intended work more concrete. Hardware design involves engineering software and artifacts from shipped hardware. Materials science covers simulated laboratory equipment and experimental data. Finance involves transaction modelling; software engineering includes debugging, migrations and longer tasks in production codebases.

The alignment category asks a different kind of question: whether agents remain honest and respect their authority under pressure. Its settings span cybersecurity, insurance and healthcare. Including that category means an overall score is not simply a measure of technical problem-solving.

The score depends on the test machinery

According to the launch methodology, tasks use a mixture of binary and partial grading. The headline result averages three attempts. Domains receive equal weight, and partners receive equal weight within domains that have several contributors. A large task category therefore does not automatically dominate the aggregate just because it contains more tasks.

The default time budget is eight hours, with shorter limits for some tasks. SemiAnalysis says it used each model’s native harness where available and OpenCode otherwise. The comparison consequently includes the software environment around a model, rather than isolating model weights under one identical interface.

The firm also says safety classifiers affect some overall scores, particularly in cybersecurity, biology and alignment. That explanation belongs to the evaluators; the reviewed material does not establish how much of every low result is refusal rather than inability.

For a buyer, these details change the useful question. A model’s overall position is a starting point for inspecting the relevant domain, task conditions and failure cases. It is not a substitute for testing the actual workflow, including its permission boundaries and review requirements. That is an implication of the benchmark’s design, not a measured deployment outcome.

The launch post plans updates every three to six months. UpThat has read the primary launch material, leaderboard and domain catalogue, but has not rerun the evaluations. A low opening ceiling gives the test room to distinguish future improvements; whether those improvements transfer to a particular workplace remains a separate question.