GDPx: even the best AI agent scores 12%

On November 2, 2026, SemiAnalysis published GDPx, a frontier agentic benchmark from Max Kan, Andrew Megalaa, and Dylan Patel. The headline they put on the page: even the best model scores just 12%.

What it covers. GDPx spans 206 tasks across 11 high-value domains: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement (RSI), hardware design, materials science, alignment, and software engineering. Domains are weighted equally. The default task time limit is 8 hours. Headline scores are pass@1 averaged over 3 attempts. Where a model has a native harness, SemiAnalysis used it (Claude Code for Opus 5.5, Codex for GPT-6 Astra, Antigravity for Gemini 4 Carbon; OpenCode otherwise).

The leaderboard (SemiAnalysis scores). Claude Opus 5.5 sits first at 12.0%. Gemini 4 Carbon is second at 10.9%. GPT-6 Astra is third at 9.6%. SemiAnalysis notes Gemini 4 Carbon is a real step for Google and the only model in the same ballpark as Opus and Astra — and that Carbon is not generally available yet.

Where everyone fails. Finance and legal tasks are so hard that models are basically at 0% under the intended binary grading (partial scores are shown for those domains). SemiAnalysis also says safety classifiers sandbag Opus 5.5 and GPT-6 Astra on cyber, biology, and alignment; excluding those three domains changes the ranking picture. Domain scores are spiky: Opus wins quant trading, hardware design, and chip design in their write-up; Carbon leads biology and edges software engineering; Astra leads RSI and materials science.

Why they built it. SemiAnalysis argues GDP-share / O*NET-style "knowledge work" benches like OpenAI's older GDPval are a poor fit for today's agents: spoon-fed prompts, thin environments, rubric noise. GDPx partners with specialized RL-environment vendors per domain so each task includes a realistic work environment (deal rooms, live services, lab sims, CAD tools, and so on).

Caveats. These are SemiAnalysis's own scores, not independently re-run by MODEL. Safety classifiers depress Opus/Astra in cyber, biology, and alignment. Finance and legal are nearly 0 for all models under intended grading. Gemini 4 Carbon is not generally available per SemiAnalysis. Not tested by MODEL. Not sponsored.

Sources: SemiAnalysis GDPx newsletter (opens in a new tab); gdpx.semianalysis.com (opens in a new tab).

MODEL Editors