SemiAnalysis launches GDPx benchmark covering 11 agent domains – Unite.AI

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

The leaderboard V1.0

The official GDPx website lists the first version with 206 activities performed with 14 specialized data partners and its ranking ranks 13 models. Claude Opus 5.5 tops the list with an average cost of $51.41 per task with 548,000 output tokens. Google’s Gemini 4 Carbon comes in second with 10.9% and $16.29 per task, while OpenAI’s GPT-6 Astra is third with 9.6% and $53.99 per task. Meta’s Muse Wonder Full is in fourth place with 9.0% with no public price listed, and Z.ai’s GLM 5.3 is in 13th with 2.7%.

SemiAnalysis describes GDPx as a true frontier benchmark precisely because the maximum score remains so low. The main GDPx score averages each model’s pass@1 result over three attempts, weights each domain equally, and applies the intended scoring method for each task, binary or continuous.

Design, partners and methodology

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

Each domain was created with specialized partners: Aria, Arcverex and Ergodex for chip design; Arithmetic for computer security; Latch Bio for biology; Housing context for finance; Chakra for legal; Third Moment for quantitative trading; Extrapolate for RSI; Beautiful hardware design machines; ValenceForge and Emulated for materials science; Fleet for alignment; and Datacurve for software engineering.

Tasks have 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 a fallback. SemiAnalysis required each partner to pass its automated QA on the first try, with only small errors at worst, and reports that the same process flags nearly 50% of tasks in the Terminal Bench 4.0 benchmark as broken.

The criticism of GDPval

The announcement positions GDPx as the successor to OpenAI’s GDPval, released on September 25, 2025, which the authors say remains the most popular general-purpose benchmark on major model release boards. They criticize GDPval’s tasks as non-interactive, with all the necessary context injected into the prompt, and cite human expert rater agreement of 71% on its gold subset. Instead, each GDPx asset is provided with a realistic environment: the 24 financial assets, for example, replicate a real private equity transaction in which a fund buys approximately 62% of a snack food manufacturer for $880 million plus an Earnout, reconstructed from approximately 4,000 anonymized files.

Domain results and stated limitations

SemiAnalysis reports Claude Opus 5.5 as the best all-round model, winning in quantitative trading, hardware design, and chip design; Gemini 4 Carbon as best biological model, with a narrow win in software engineering; and GPT-6 Astra as winner in CSR and materials science. He warns that Opus 5.5 and GPT-6 Astra’s overall scores are held up by their security classifiers in the cyber, biology, and alignment domains, and describes Meta’s Muse Wonder Full as the best cyber model available by a significant margin.

Financial and legal tasks are so difficult that each model scores zero against the expected binary classification, so SemiAnalysis reports average subscores for those domains. In quantitative trading, 70 out of 111 graded executions performed worse than the reference strategy the agent was asked to beat, and no model solved the task of winsorizing the public aapl target. In fleet alignment computer tasks, 196 out of 200 races broke at least one behavior rule and 173 broke a rule rated as severely misaligned; in healthcare activities, 19 out of 100 rides scored above zero. Regarding costs, SemiAnalysis places GPT-6 Luna and DeepSeek at the low end of the cost-performance frontier, Opus at the high end, and OpenAI in the middle.

SemiAnalysis says GDPx is intended as a living benchmark, with new releases planned every 3-6 months, and lists [email protected] for data companies interested in participating and labs seeking evaluation of future releases.