The GDPx benchmark has launched with Claude Opus 5.5 leading at 12.0%, ahead of Gemini 4 Carbon at 10.9% and GPT-6 Astra at 9.6%. SemiAnalysis released the agent evaluation on November 2, 2026, covering 206 tasks in 11 specialist domains.

The low scores are striking, but they require context. GDPx emphasizes long, tool-using assignments built with specialist data partners, not short questions or a representative sample of all professional work. Its debut results measure performance under one demanding methodology rather than the percentage of jobs an AI system can perform.

The first GDPx leaderboard establishes four headline results:

• Claude Opus 5.5 ranks first at 12.0%.
• Gemini 4 Carbon follows at 10.9%.
• GPT-6 Astra places third at 9.6%.
• All tested systems remain below 13% overall.

GDPx Benchmark Covers 206 Tasks Across 11 Domains

The official GDPx leaderboard describes a collaboration with 14 specialized data partners. Its domains include chip design, cybersecurity, biology, finance, law, quantitative trading, recursive self-improvement, hardware design, materials science, alignment and software engineering.

The distribution is uneven because each field requires different evidence and tooling. Chip design contributes 34 tasks, finance 24, materials science 23 and hardware design 21. Cybersecurity and alignment each contribute 18, while software engineering has 15 and biology has 13.

SemiAnalysis says the environments are intended to resemble specialist workflows. Finance tasks use roughly 4,000 anonymized files connected to a reconstructed private-equity transaction. Cybersecurity tasks run inside live multi-service environments, while hardware assignments use computer-aided design and programmable logic tools.

That design makes the benchmark closer to an agent test than a conventional question set. A model must inspect data, operate tools, manage intermediate steps and produce an artifact that can be graded. The benchmark gives tasks a default limit of eight hours, allowing approaches that extend far beyond a single response.

GDPx is also selective. Its authors intentionally focus on fields where they expect frontier AI to generate high economic returns, excluding areas such as government, real estate and insurance. The resulting score should therefore be read as performance across the chosen high-value domains, not a national productivity measure.

Claude Opus 5.5 Leads a Closely Packed Top Three

Claude Opus 5.5's 12.0% result gives Anthropic the initial lead. Google’s Gemini 4 Carbon is 1.1 percentage points behind, while OpenAI’s GPT-6 Astra trails Carbon by 1.3 points. Meta’s Muse Wonder Full ranks fourth at 9.0%, according to the published table.

The middle of the ranking is tightly grouped. Claude Fable 5.1 scores 8.5%, GPT 6.1 Sol reaches 8.4% and Claude Sonnet 5.5 records 8.2%. Gemini 4 Argon follows at 7.3%, while Muse Wonder scores 6.8%.

Open-weight systems appear lower in the debut comparison. DeepSeek 4.1 Flash 1017 reaches 5.4%, and Z.ai’s GLM 5.3 records 2.7%. OpenAI’s lower-cost GPT-6 Luna scores 5.1%, while xAI’s Grok 4.7 reaches 3.9%.

The ranking changes by subject. SemiAnalysis reports that Opus leads quantitative trading, hardware design and chip design. Carbon leads biology and narrowly takes software engineering, while Astra performs best in recursive self-improvement and materials science. No single model dominates every kind of specialist work.

Safety systems also complicate comparisons. SemiAnalysis says classifiers reduced results for Opus and Astra in cybersecurity, biology and alignment. That does not make the restrictions irrelevant: deployed agents must balance usefulness with policies governing dangerous or sensitive work. It does mean the aggregate score combines capability and refusal behavior.

GDPx Scoring Rewards Completed Work, Not Plausible Answers

For its headline metric, GDPx reports pass@1 averaged across three attempts. Tasks use binary or continuous grading depending on the domain, and every domain receives equal weight. Data partners within a domain are also weighted equally, preventing the largest task collection from automatically controlling the overall ranking.

Models run through their native agent harness when one is available. The methodology uses Claude Code for Opus, Codex for Astra and Antigravity for Carbon, with OpenCode used otherwise. That decision reflects how buyers encounter agent products, but it means results compare model-and-tool systems rather than isolated language models.

Finance and legal illustrate the difficulty of binary evaluation. SemiAnalysis says every model received zero under the intended pass-or-fail scoring in those domains, even though partial-credit measures separated their work. A zero can therefore mean an output missed a completion threshold, not that it contained no useful analysis.

Average cost adds another dimension. The leaderboard lists about $51.41 per task for Opus, $16.29 for Carbon and $53.99 for Astra. GPT 6.1 Sol is listed at $9.80, while DeepSeek 4.1 Flash and GPT-6 Luna are below $1. These are benchmark-run averages, not general subscription prices.

GDPx Results Need Reproduction and Version Tracking

The debut provides a useful snapshot because its tasks require sustained work in realistic environments. Independent benchmark trackers and technology publications have confirmed the published rankings, but the underlying results remain the benchmark operator’s measurements. They have not yet been broadly reproduced by outside laboratories.

The selected domains, partner-built environments and equal-weighting rule all reflect design judgments. A different mix of fields, grading thresholds or agent harnesses could produce another order. Future model updates could also change refusal patterns or tool reliability without altering the model names shown on a static chart.

The most valuable next disclosures would include reproducible task subsets, error categories and versioned results as models and harnesses change. SemiAnalysis says GDPx is intended to evolve, so comparisons should record the benchmark version and date rather than treating 12.0% as a permanent model characteristic.

For now, GDPx’s strongest finding is not that AI agents can complete exactly 12% of professional work. It is that even leading systems finish only a small share of this deliberately difficult, tool-intensive task set, with material differences across domains, costs and safety behavior.

Illustrative photo: laboratory researchers using computer equipment, photographed by Chidera Faustina Okeke on Unsplash. Licensed under the Unsplash License. Cropped to 16:9; the image does not show GDPx testing.

Related Research

• Google Unveils Gemini 4 Carbon With 88.2% Vulnerability Score
• GLM-5.3 Cyberattack Tests Show 3 of 13 Simulations Completed