Skip to content
Neural Daily

ResearchNews

GDPx: top AI agent scores 12% on SemiAnalysis benchmark

🇭🇺 Olvasd el magyarul →

SemiAnalysis has launched GDPx, a benchmark that tests frontier models in realistic work environments across 11 high-value domains. The leader, Claude Opus 5.5, scores just 12%, and on finance and legal tasks virtually every model scores zero.

By: Mikola Máté Patrik · · 3 min read

GDPx benchmark website by SemiAnalysis showing the leaderboard with Claude Opus 5.5 on top
Image: SemiAnalysis · Screenshot · Screenshot of the official GDPx website · Source

What is GDPx?

SemiAnalysis argues that OpenAI's GDPval, still reported on most model cards, is a pre-agent benchmark. Its tasks are tightly scoped, the model gets all context in the prompt, and the output is usually a single generated document. GDPx instead drops the agent into a simulated work environment with a default time limit of 8 hours per task.

The benchmark has 206 tasks across 11 domains, built by 14 specialized RL environment startups: recursive self-improvement (RSI), hardware design, cybersecurity, biology, finance, legal, quant trading, chip design, software engineering, materials science and alignment. Government, real estate and insurance were deliberately left out. SemiAnalysis expects the first trillion dollars of AI revenue to come from sectors where frontier models deliver high ROI.

What the tasks look like

  • Finance: a deal room of about 4,000 files built from a real private equity transaction, where the agent has to gather context itself.
  • Cybersecurity: live environments built on real, undisclosed zero-days with multiple running services.
  • Hardware design: SolidWorks, Siemens NX or Fusion 360 sessions where the grader perturbs features to check that the design rebuilds parametrically.
  • RSI: diagnosing a degraded training run from about 1,000 RL rollouts (21.2 million tokens).

Models ran in their native harness where available (Claude Code, Codex, Antigravity), others in OpenCode. Scores average three attempts, with all domains weighted equally.

The results

Claude Opus 5.5 leads the overall board with 12.0%, followed by Gemini 4 Carbon (10.9%) and GPT-6 Astra (9.6%). The best open-weight model is DeepSeek 4.1 Flash (5.4%), while GLM 5.3 scores 2.7%. Costs vary widely: a task costs about $51 with Opus, $16 with Carbon and under $1 with DeepSeek 4.1 Flash.

GDPx leaderboard across all 11 domains: Claude Opus 5.5 12.0%, Gemini 4 Carbon 10.9%, GPT-6 Astra 9.6%
SemiAnalysis · Screenshot · Leaderboard chart from the SemiAnalysis GDPx post · Source

Source: SemiAnalysis (GDPx). The full leaderboard across all 11 domains, averaged over 3 attempts, with cost per task.

Safety classifiers change the picture. Excluding cybersecurity, biology and alignment, where some models' safeguards often step in, GPT-6 Astra takes the lead with 11.8%, just ahead of Opus 5.5 at 11.7%. SemiAnalysis says this makes Muse Wonder Full the best available cyber model today, and Gemini 4 Carbon likely wins biology because of weaker classifiers.

What to take from it

Models are spiky. Opus wins quant trading and chip design, Astra wins RSI and materials science, and Carbon edges out software engineering. On finance, every model scores 0% under binary grading, so partial scores are reported instead. SemiAnalysis's own QA process flagged nearly half of Terminal Bench 4.0 tasks as broken, which shows how hard it is to build a fair agent benchmark today.

The practical lesson: when picking an agent for a domain, look at results filtered to that domain rather than the headline ranking. Factor in safety classifiers too, since a strong model is of little use if its safeguards stop the work.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles