The best AI agent solves 1 in 8 expert tasks. Here is what that means for builders.

A research firm called SemiAnalysis just put the world's best AI models through 206 real expert tasks — chip design, finance, law, biology, quant trading, and more. The top model, Claude Opus 5.5 from Anthropic, finished just 12% of them. Not because AI is bad. Because the tasks are genuinely hard. AI agents still fail at 88% of real expert work, and that number is actually a useful map for anyone building with AI today.

The breakdown matters more than the headline score. Claude led on quant trading and chip design. Google's Gemini 4 Carbon did best in biology and software engineering. Finance and legal? Basically every model scored zero — the tasks were too hard for any of them to crack a single one. That is not a blanket failure of AI. It is a signal: those fields are not ready for autonomous agents handling the hard judgment calls.

Most AI tests measure whether a model knows the right answer. GDPx asked whether it can actually do the work — write real code, run real simulations, finish multi-step tasks where specialists check the result. Every domain was built with a partner company whose job was to verify the output. That is what makes these numbers more trustworthy than the benchmarks where 50% of tasks turn out to be broken, as SemiAnalysis found when it audited other popular tests.

For anyone building products or automations with AI right now, this benchmark is a map, not a verdict. Quant trading and chip design are moving fast. Finance and law are not ready for autonomous agent work — at least not for anything requiring genuine expert judgment. The practical takeaway: focus agent work where models are already getting real results, not where they score zero. SemiAnalysis plans to refresh GDPx every three to six months, so the map will keep updating as the models improve.