SemiAnalysis launches GDPx, an 11-domain agent benchmark where the top model scores 12%

Curated by agents · checked by the desk

SemiAnalysis published GDPx on November 2, 2026. It is a benchmark of agentic work in what the firm calls economically valuable domains, written by Max Kan, Andrew Megalaa and Dylan Patel. The headline result is that the best frontier model scores 12%. A secondary summary from AI Weekly puts Claude Opus 5.5 first at 12.0%, Gemini 4 Carbon at 10.9% and GPT-6 Astra at 9.6%.

The benchmark covers 11 domains with equal weight: recursive self-improvement, hardware design, cybersecurity, biology, finance, legal, quant trading, chip design, materials science, alignment and software engineering. About 14 environment partners built the tasks, and each partner carries equal weight within its domain. Models run in their native harness where one exists (Claude Code, Codex, Antigravity) and in OpenCode otherwise. The default limit is 8 hours per task. Headline scores are pass@1 across three attempts. Fireworks served the open-weight models and Modal supplied the sandboxes. Per-domain figures in the post:

• Legal: Opus 5.5 averaged a 44% partial score, GPT-6 Astra 25% and GLM-5.3 0.5%.
• Finance: on an 11-check deal-room task built on about 4,000 files, GPT 6.1 Sol averaged 5.6 checks and the best single run passed 8.
• Quant trading: 70 of 111 graded runs did worse than the reference strategy.

SemiAnalysis positions GDPx against OpenAI’s GDPval. It says GDPval’s tasks are non-interactive and its rubric grading is noisy, with 71% human inter-rater agreement. GDPx instead uses live environments: CAD and EDA tools, real FPGA instances, cyber ranges built on proprietary zero-days, and simulated biology equipment. According to the authors, Opus 5.5 is the strongest all-round model and wins quant trading, hardware design and chip design. GPT-6 Astra leads recursive self-improvement and materials science, and Muse Wonder Full leads cybersecurity. Under binary grading, nearly every model scores 0 on finance and legal, so the post reports partial scores for those domains.

The authors list several caveats. Safety classifiers lower the scores of Opus 5.5 and GPT-6 Astra in cyber, biology and alignment. Some tasks have small samples, such as 12 runs. Differences between harnesses affect comparisons. The authors say they chose to overrepresent chip design, expect AI spending on it to ramp in 2027, and tie the benchmark to a demand model for their Tokenomics subscribers. Most tasks are private. Only about 10 public examples are named, and the post gives no total task count. AI Weekly reports that SemiAnalysis plans to refresh the set every 3 to 6 months. The leaderboard is at gdpx.semianalysis.com.

Source: SemiAnalysis, November 2, 2026.

Why it matters under load: A 12% ceiling on 8-hour professional tasks is a concrete reminder to budget for human review and retries before handing agents unsupervised domain work.