SemiAnalysis 发布 GDPx 基准，覆盖 11 个智能体领域

SemiAnalysis on November 2, 2026 released GDPx, a benchmark for general agentic capabilities spanning 11 professional domains, in an announcement authored by Max Kan, Andrew Megalaa, and Dylan Patel. Anthropic’s Claude Opus 5.5 leads the debut leaderboard with a score of 12.0%.

The V1.0 Leaderboard

The official GDPx site lists the first version at 206 tasks built with 14 specialized data partners, and its leaderboard ranks 13 models. Claude Opus 5.5 tops the ranking at an average cost of $51.41 per task with 548K output tokens. Google’s Gemini 4 Carbon places second at 10.9% and $16.29 per task, while OpenAI’s GPT-6 Astra is third at 9.6% and $53.99 per task. Meta’s Muse Wonder Full ranks fourth at 9.0% with no public price listed, and Z.ai’s GLM 5.3 is 13th at 2.7%.

SemiAnalysis describes GDPx as a true frontier-level benchmark precisely because the top score remains so low. The headline GDPx score averages each model’s pass@1 result across three attempts, weights every domain equally, and applies each task’s intended grading method, binary or continuous.

Design, Partners, and Methodology

The domains are chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement (RSI), hardware design, materials science, alignment, and software engineering; SemiAnalysis says it intentionally forgoes industries such as government, real estate, and insurance. The authors argue that the first trillion dollars of AI revenue will concentrate in sectors that can deploy frontier intelligence at high ROI, and that building a general benchmark around share of GDP or the SOC/O*NET occupational taxonomy is a mistake.

Each domain was built with specialized partners: Aria, Arcverex, and Ergodex for chip design; Arithmetic for cybersecurity; Latch Bio for biology; Living Context for finance; Chakra for legal; Third Moment for quant trading; Extrapolate for RSI; Beautiful Machines for hardware design; ValenceForge and Emulated for materials science; Fleet for alignment; and Datacurve for software engineering.

Tasks carry a default time limit of eight hours, and models run in their native harness when available — Claude Code for Opus 5.5, Codex for GPT-6 Astra, and Antigravity for Gemini 4 Carbon — with OpenCode as the fallback. SemiAnalysis required every partner to pass its automated QA on the first try with only minor mistakes at worst, and reports that the same process flags nearly 50% of tasks in the Terminal Bench 4.0 benchmark as broken.

The GDPval Critique

The announcement positions GDPx as a successor to OpenAI’s GDPval, released September 25, 2025, which the authors say remains the most popular general-purpose benchmark on major model release cards. They criticize GDPval’s tasks as non-interactive, with all necessary context spoon-fed in the prompt, and cite human expert inter-rater agreement of 71% on its gold subset. Every GDPx task instead ships with a realistic environment: the 24 finance tasks, for example, replicate a real private-equity transaction in which a fund buys about 62% of a snack manufacturer for $880 million plus an earnout, reconstructed across roughly 4,000 anonymized files.

Domain Results and Stated Limitations

SemiAnalysis reports Claude Opus 5.5 as the best all-around model, winning quant trading, hardware design, and chip design; Gemini 4 Carbon as the best biology model, with a narrow win in software engineering; and GPT-6 Astra as the winner in RSI and materials science. It cautions that the Opus 5.5 and GPT-6 Astra overall scores are sandbagged by their safety classifiers in the cyber, biology, and alignment domains, and describes Meta’s Muse Wonder Full as the best available cyber model by a meaningful margin.

The finance and legal tasks are difficult enough that every model scores zero under the intended binary grading, so SemiAnalysis reports average partial scores for those domains. In quant trading, 70 of 111 graded runs did worse than the reference strategy the agent was asked to beat, and no model solved the public aapl-target-winsorization task. In Fleet’s alignment cyber tasks, 196 of 200 runs broke at least one behavior rule and 173 broke a rule rated severely misaligned; in the healthcare tasks, 19 of 100 runs scored above zero. On cost, SemiAnalysis places GPT-6 Luna and DeepSeek at the low end of the cost-versus-performance frontier, Opus at the high end, and OpenAI in the middle.

SemiAnalysis says GDPx is intended as a living benchmark, with new versions planned every 3–6 months, and lists gdpx@semianalysis.com for data companies interested in participating and labs seeking evaluation of future releases.