GDPx: the best AI agent finishes only 12% of 206 hard, high-value work tasks

SemiAnalysis's new GDPx benchmark gives agents 206 long, real-world tasks from 11 fields (including chip design, cyber security, biology, finance, legal and quant trading), with up to eight hours per task. Claude Opus 5.5 came top with 12.0% (about $51 per task), ahead of Gemini 4 Carbon (10.9%, not generally available) and GPT-6 Astra (9.6%); DeepSeek 4.1 Flash scored 5.4% for under $1 a task. Finance and legal scored close to zero under pass-or-fail grading, and most tasks are private, so nobody else can rerun them yet. On the same day OpenAI published LifeVal, a benchmark of everyday personal-assistant jobs (childcare, government services, money, health and shopping) built from surveys of thousands of people; it shares only selected scores, with no overall score or task set.

Why it mattersA useful reality check: today's agents handle chunks of work, not whole expert jobs. Give them well-defined steps with a person checking the result, and judge vendors on tasks like yours rather than on their own benchmarks.

Curated by Sara Simeone as part of our human editorial coverage. The original sources are linked below; curation is distinct from a Hype or Fact? assessment.

How we knowCompany figures are labelled as their own claims; anything we couldn't verify is labelled as a report or a claim.