SemiAnalysis's new GDPx test finds the best AI model succeeds only 12% of the time

A tough test with a low top score

Research firm SemiAnalysis has released a benchmark called GDPx, and the first results are humbling. As Crypto Briefing reported, the best-performing model, Anthropic's Claude Opus 5.5, reached an overall success rate of 12.0% on the GDPx v1.0 leaderboard.

Google's Gemini 4 Carbon came second at 10.9, and OpenAI's GPT-6 Astra third at 9.6. Even the winner fails most of what it is asked to do, and the top three sit within a few points of each other.

SemiAnalysis unveiled the test on November 2, 2026. It describes GDPx as an attempt to build the best available yardstick for general agentic capabilities. 'Agentic' here means AI that carries out several steps toward a goal instead of just answering a single question.

Why these 11 fields were chosen

GDPx tests frontier models (the most capable systems from the leading labs) across 11 domains. Chip design, biology and finance are among them. The choice is deliberate and rests on a thesis from SemiAnalysis: the first trillion dollars of AI revenue will gather in sectors that can use top-level intelligence for a high return on investment, not spread evenly across the economy.

That reasoning shaped the design. The firm did not weight tasks by each industry's share of GDP, and it did not follow the SOC/O*NET system, an official classification of jobs across the workforce. It picked the areas where it expects frontier models to matter most.

The trade-off is that some sectors are missing on purpose. Government and insurance are excluded because SemiAnalysis sees lower potential return from AI there. Readers should keep this in mind: GDPx is a measure of one view of where AI value will land, not a map of the whole economy.

Who built the tasks

To write the challenges, SemiAnalysis worked with 14 specialized startups. These companies provide data and reinforcement learning services; reinforcement learning is a training method in which a model improves by being scored on its attempts. Each partner helped create tasks inside particular domains of the framework.

Using specialists makes sense for fields like biology or chip design, where a generalist might write tasks that are too easy or miss the real problems. The source does not describe how tasks are scored in detail, so we cannot say more about how the 12.0% is calculated beyond calling it an overall success rate.

No single winner across the board

The overall ranking hides a more interesting pattern. Gemini 4 Carbon took first place in both biology and software engineering. GPT-6 Astra won in recursive self-improvement and in materials science. Claude Opus 5.5 holds the top overall spot without winning everything.

The research findings state that no current model solves tasks comprehensively across all the evaluated domains. A model that looks strong in one specialty can look average in another.

That complicates any simple story about which AI lab is ahead. The headline ranking favors Anthropic, but Google and OpenAI each show clear strengths in particular technical areas.

What happens next

SemiAnalysis says GDPx will not stay fixed. The firm plans to update it every 3 to 6 months, folding in feedback and adding more partners. Scores on version 1.0 are therefore a snapshot, and later versions may not be directly comparable.

The findings suggest that models clearing roughly one in ten tasks shows capabilities are advancing, but large barriers remain before any model reaches broad skill in high-stakes professional work. For now, the sensible reading is modest: progress is real, and so are the gaps.