SemiAnalysis launches GDPx agent benchmark; Claude Opus 5.5 leads with just 12% of 206 tasks

What SemiAnalysis released

SemiAnalysis, the semiconductor and AI research firm led by Dylan Patel, published GDPx on 2 November 2026, describing it as a benchmark for measuring agentic capabilities across the most valuable domains of the economy. The launch post, written by Max Kan, Andrew Megalaa and Dylan Patel, comes with a public leaderboard and data explorer at gdpx.semianalysis.com.

Version 1.0 has 206 tasks built with 14 specialist data partners, spread over 11 domains: chip design, cybersecurity, biology, finance, legal, quantitative trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. Chip design has the most tasks, 34, followed by finance with 24 and materials science with 23, according to the site and several write-ups. Tasks ship with simulated working environments instead of a single prompt; the 24 finance tasks, for example, rebuild a private equity deal across roughly 4,000 anonymized files that the agent has to search and work through.

SemiAnalysis says it picked domains where frontier models could earn the most money, arguing on its launch thread that the first trillion dollars of AI revenue will be concentrated in sectors that can deploy frontier intelligence at high return, not spread across the whole economy. That is why areas such as government, real estate and insurance are left out.

The first leaderboard

The headline score averages each model's pass@1 result over three attempts per task, weights the 11 domains equally and applies each task's own grading method, binary or continuous. On that measure Claude Opus 5.5 from Anthropic ranks first at 12.0%. Gemini 4 Carbon from Google DeepMind is second at 10.9% and GPT-6 Astra from OpenAI third at 9.6%, according to the leaderboard as reported by Unite.AI, Crypto Briefing and other outlets.

Two caveats come from SemiAnalysis itself. Finance and legal tasks are hard enough that every model scores zero under the intended binary grading, so the firm reports average partial scores for those two domains. And some models, Opus 5.5 and GPT-6 Astra among them, have lower scores than they might otherwise reach because of their safety classifiers, which SemiAnalysis says sandbagged their overall results. The site also charts cost per task for most models, so buyers can weigh price against success rate.

The top three are the newest frontier models from the three largest labs. Anthropic released Opus 5.5 on 22 September 2026 as the first model of its Claude 5.5 family, with a 1 million token context window. OpenAI released GPT-6 Astra on 3 September as the most capable GPT-6 model. Google DeepMind announced Gemini 4 Carbon only on 28 October and is phasing it in, with cyber defenders getting access first, so the GDPx result is one of the first outside measurements of that model. The gap between first and third place is 2.4 percentage points, small next to the distance all three still have to cover.

Why it matters and how it differs from GDPval

GDPx positions itself as a harder successor to GDPval, the benchmark OpenAI introduced in September 2025 to grade models on work products from dozens of occupations. GDPval and its derivatives are still the economic benchmarks most often cited on model release cards. The SemiAnalysis authors argue that GDPval tasks are not interactive, hand the model all context in the prompt, and show only 71% agreement between human expert graders on its gold subset, so high scores say little about whether an agent can do real, multi-step work.

The low numbers are the point. Many widely used tests are close to saturated, which leaves little room to tell frontier models apart. A benchmark where the leader scores 12% gives room to track progress over the next generations and, because it is split by domain, shows which industries may see working agents first. It joins other recent attempts to measure frontier models with private or harder task sets, such as the Artificial Analysis Intelligence Index v5. More coverage sits on the AI models hub.

What to watch next

SemiAnalysis is a commercial research firm that sells analysis to the industry it measures, and its tasks were built with partners that sell data and environments to AI labs, so independent checks of the task set and grading will matter. Open points include how often the task set is refreshed, whether labs will report GDPx scores on their own model cards, and how much of the gap between first and third place survives once safety classifier refusals are accounted for.

The per-domain results are the part most likely to be used outside the research community. A model that does well on chip design or software engineering but scores zero on legal work tells companies in those fields something different from a single headline number. Readers who want to compare models across more benchmarks can follow the model pages linked above, which keep a dated timeline of every result covered on this site.

Questions people ask

Which model leads the GDPx benchmark?

Anthropic's Claude Opus 5.5, with 12.0% on the first GDPx leaderboard published on 2 November 2026, ahead of Gemini 4 Carbon at 10.9% and GPT-6 Astra at 9.6%.

What does GDPx measure?

How well AI agents complete 206 interactive tasks in 11 high-value domains, including chip design, cybersecurity, biology, finance, legal and quantitative trading, with each domain weighted equally.

Who made GDPx?

SemiAnalysis, the research firm led by Dylan Patel, with 14 data partners. The launch post is by Max Kan, Andrew Megalaa and Dylan Patel.

Sources

1. GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains (SemiAnalysis, 2 Nov 2026)
2. GDPx: The General Agentic Benchmark (leaderboard) (SemiAnalysis, 2 Nov 2026)
3. SemiAnalysis Launches GDPx Benchmark Spanning 11 Agentic Domains (Unite.AI, 2 Nov 2026)
4. SemiAnalysis unveils GDPx benchmark, and the best AI model scores 12% (Crypto Briefing, 2 Nov 2026)
5. SemiAnalysis launch thread on X (SemiAnalysis, 2 Nov 2026)