Best AI solves 12% of 206 expert tasks

The best AI model solved just 12 percent of a new test of expert work, research firm SemiAnalysis said on the 2nd of November 2026. The test, called GDPx, has 206 tasks in 11 fields.

Anthropic's Claude Opus 5.5 came first. Google's Gemini 4 Carbon scored 10.9 and OpenAI's GPT-6 Astra scored 9.6.

What the test asks

The fields include chip design, law, finance, biology and software, SemiAnalysis said. Each task can run up to eight hours.

Every model works as an agent, a program that takes many steps alone. Each used its maker's own tool where one exists, such as Claude Code for Opus.

Finance and law scored zero

The finance and legal tasks were so hard that basically every model scored 0, SemiAnalysis wrote. One finance task used about 4,000 files from a real company sale.

On a legal task, no model wrote a usable contract markup in 20 tries. Opus earned the most partial credit, at 44 percent.

Different models win different fields

Opus won quant trading, hardware design and chip design. Gemini 4 Carbon won biology and software, and GPT-6 Astra won materials science.

SemiAnalysis said safety filters pulled down some scores in cyber, biology and alignment. Meta's Muse Wonder Full was the best cyber model.

Below we break down the scores

Opus leads Gemini 4 Carbon by 1.1 points. If every task counted the same, 12 percent of 206 would be about 25 tasks, by our arithmetic.

Opus cost about 51 dollars a task, against about 16 dollars for Carbon, Dealroom reported. That makes Opus roughly three times as costly.

What is confirmed and what is not

The scores are SemiAnalysis's own runs, and it chose the fields. No outside group has checked them yet.

Key facts

SemiAnalysis released GDPx with 206 tasks on the 2nd of November 2026.

Claude Opus 5.5 led with 12 percent.

Nearly every model scored 0 on finance and legal tasks.

SemiAnalysis ran the tests itself.