GDPx benchmark: the best AI agent scores 12%

A new agent benchmark just reset expectations. SemiAnalysis released GDPx, a test of AI agents doing real professional work inside realistic environments, and the best model scores 12%. Claude Opus 5.5 leads, Gemini 4 Carbon scores 10.9 and GPT-6 Astra 9.6, per the GDPx leaderboard. On the finance and legal tasks, SemiAnalysis says "basically every model scores exactly 0."

What actually happened

From SemiAnalysis's launch post and the leaderboard site:

• Scope: version 1.0 has 206 tasks across 11 domains: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. Government, real estate and insurance are left out on purpose.
• Built with specialists: SemiAnalysis worked with 14 data partners, mostly RL-environment startups, one or more per domain.
• Real environments: this is the key difference from OpenAI's GDPval, which SemiAnalysis says feeds the model all the context in the prompt. A GDPx finance task drops the agent into a deal room of about 4,000 anonymized files from a real private-equity transaction, and the agent must find the right documents itself.
• Safety classifiers matter: SemiAnalysis says Opus 5.5 and GPT-6 Astra scores are pulled down in cyber, bio and alignment, where safety classifiers sometimes stop the model.
• Spiky results: Opus 5.5 wins quant trading, hardware and chip design. Gemini 4 Carbon edges out software engineering. GPT-6 Astra wins RSI and materials science.

Why it matters for your business

Vendor demos show agents on clean, well-scoped tasks. GDPx shows what happens when the agent has to dig through a messy folder to find the facts first. That is what your business looks like: a shared drive, three versions of the price list, and an inbox with the real answer in a reply thread.

The lesson is not "agents are useless." It is that the hard part of most work is gathering context, and that is where agents fail. So design for it. Do not hand an agent a whole drive and a vague goal. Give it a narrow job, a curated set of sources, and a clear definition of done. Put a human check on anything that touches money or a contract.

Also note how spiky the results are. No model wins every domain. The model that tops a public leaderboard may not be the best one for your quotes, your support tickets or your catalog. The only benchmark that answers that question is one built from your own tasks, with your own files, graded against what your best employee would have done.

Sources: SemiAnalysis: GDPx launch post, GDPx leaderboard.