Just 12 per cent on 206 expert tasks: what the GDPx benchmark reveals about the real limits of today's AI agents

A new benchmark built from real, paid expert work has tested twenty leading AI models, and the best of them scored just 12 per cent.

This week, the research firm SemiAnalysis released a benchmark called GDPx. It contains 206 tasks across 11 industries, chosen because people do them for a living: chip design, cybersecurity, legal work, quantitative trading, biology, software engineering and more. Fourteen industry partners supplied the tasks, and twenty leading models were evaluated. The benchmark is notable not for its difficulty but for its raw material: work that people are paid to do.

The top-scoring model, Claude Opus 5.5, achieved only 12 per cent. Gemini 4 Carbon scored 10.9, and GPT-6 Astra 9.6.

Yet the leaderboard is a starting point, not a verdict. GDPx uses one firm's tasks and its own evaluation, has not been independently verified, and that firm sells research to the industry it evaluates. Treat the numbers with caution - but the direction is right, because it measures models on paying work rather than on material designed to attract them.

The top scorer spent about $51 per completed task on average; the runner-up, within a point on the leaderboard, spent only about $16. If your agent strategy assumes capability and cost rise together, this table says otherwise.

The takeaway for leaders evaluating AI agents this quarter: every demo will claim to be ready for expert work, but a public, named benchmark has just put the ceiling at 12 per cent. That does not make agents useless; it means today's honest revenue surface is narrow, closely watched tasks with human review, not the autonomous departments sold from the stage.

So the era of the demo is over. Ask every AI vendor one question: on an independent, task-based benchmark of work like mine, show me your score and your cost per completed task. A 12 per cent ceiling is no reason to walk away from AI agents; it is a reason to buy them with evidence instead of enthusiasm.