GDPx

GDPx is an AI benchmark published by SemiAnalysis on November 2, 2026 to evaluate agents working in simulated professional environments. Version 1.0 contains 206 tasks across 11 domains, contributed by 14 specialist partners. It evaluates work involving tools, files and software environments, rather than only a model's answer to a standalone question.^[1][2][3]

The benchmark's domain selection reflects SemiAnalysis's view of where demand for frontier AI will grow. That commercial thesis motivates the selection; GDPx scores do not measure a model's contribution to gross domestic product or establish the economic value of an individual deployment.^[2]

Domains and task environments

The version 1.0 domain index lists the following distribution. Counts describe tasks, not the weight each domain receives in the overall score.^[3]

Environments differ by discipline. The finance environment reconstructs an anonymized private-equity transaction using approximately 4,000 files; each assignment supplies the deal-room material and the relevant participant's folder. Hardware tasks use engineering applications and existing design artifacts. Materials-science tasks expose simulated laboratory equipment and experimental data.^[2][3]

The alignment domain examines whether agents remain honest and within their authority under task pressure. It includes cyber, insurance and healthcare scenarios, despite insurance not being a separate headline domain. The tasks assess behavior as well as whether an agent obtains the requested outcome.^[2][3]

The domain called recursive self-improvement tests research-engineering workflows, including diagnosis of failed training runs, data quality and AI-system optimization. A result on these bounded tasks does not demonstrate an agent independently building an entire improved successor model.^[2][3]

Evaluation procedure

SemiAnalysis describes a default limit of eight hours per task, with shorter limits for some tasks. Models use their native agent harness where one is available: examples include Claude Code, Codex and Antigravity. Other models use OpenCode. The resulting measurements therefore concern model-and-harness configurations, rather than a comparison that holds the surrounding software constant.^[2]

The developers screened potential partners' tasks using automated quality checks and manual trajectory review. Their stated checks include consistency, completeness, cheating opportunities and unfair verifiers. The launch article describes requiring tasks to pass the first automated review with, at most, minor errors, while allowing partners to contest a flagged issue. These are the benchmark developers' quality-control procedures, rather than independent validation of every task.^[2]

Scoring

The official leaderboard says tasks use binary or continuous grading according to their domain. Its default score uses the intended grading method for each task and averages results across three attempts. SemiAnalysis describes the headline statistic as pass@1 across those attempts.^[1][2]

Every domain has equal weight in the overall GDPx score. Where a domain has several partners, those partners receive equal weight within it. Thus, the 34 chip-design tasks do not collectively receive more domain weight than the ten quantitative-trading tasks.^[1][2][3]

Because some tasks award continuous credit, a 12% headline score should not be interpreted as simply completing 12% of the 206 tasks. The site also exposes scoring and domain controls, which need to remain attached to any reported result.^[1]

Initial results

The following are selected results displayed by the official version 1.0 leaderboard on November 4, 2026, using its default score over all 11 domains. They are SemiAnalysis's evaluations, not an independently reproduced ranking.^[1]

These costs accompany the tested runs. They are not a model's per-token API price or a quote for a future task. The site omits Muse Wonder Full from its cost-performance plot because it has no public price.^[1]

Excluding safety-affected domains

SemiAnalysis reports that safety classifiers affect some models' results in biology, cybersecurity and alignment. The leaderboard offers an explicit filter that excludes those three domains. The remaining eight-domain comparison changes the order and scores:^[1][2]

The filtered column is a different task distribution, not a replacement for the overall score. The publisher's explanation also should not turn a safety refusal into an assumed successful solution: the observed run still receives its recorded result.^[1][2]

Released traces and reproducibility

Each partner released rollouts for one task. The public viewer offers 168 transcripts, showing agent messages, tool calls and results for those selected examples. The rest of the rollouts remain private, which SemiAnalysis says is intended to prevent repeated optimization against the test set.^[4]

These traces permit inspection of particular successes and failures. They do not provide the complete private task collection or every run underlying the aggregate leaderboard. Reproducing the full evaluation consequently requires access beyond the public transcript examples.^[1][4]

Relationship to GDPval

GDPx is distinct from OpenAI's GDPval. GDPval's first version covers 44 occupations in nine industries selected using U.S. economic and employment data. It evaluates work products supplied with reference material and uses a one-shot setup. OpenAI explicitly identifies the absence of iterative context building and multiple drafts as a limitation of that version.^[5]

GDPx instead chooses domains according to SemiAnalysis's expectations about demand for frontier intelligence and supplies interactive environments. Their task distributions, execution setups and grading differ, so a percentage on one benchmark cannot be read as the same measure on the other.^[1][2][5]

Continue reading

• SemiAnalysisSemiAnalysis is an independent research and analysis firm covering the semiconductor and artificial intelligence industries.Mentioned 8 times above
• BBenchmark (AI)In artificial intelligence and machine learning, a benchmark is a specified evaluation used to compare systems under common conditions.Mentioned 4 times above
• GGDPvalGDPval is a benchmark released by OpenAI on September 25, 2025 that measures how well frontier AI models can produce the actual deliverables of professional knowledge workMentioned 3 times above
• Harness (AI)Harness, in artificial intelligence, is the software scaffolding that wraps a machine learning model, most often a large language modelMentioned twice above
• Claude CodeClaude Code is an agentic software development product made by Anthropic. It uses models from the Claude family to inspect software projects, propose and apply changes, run development tools, and report…Also in AI Agents
• opencode (SST)opencode is an open-source AI coding agent built for the terminal. Developed by Anomaly Innovations, formerly known as the SST (Serverless Stack) team, it provides a terminal user interface (TUI) through which…Also in AI Agents

More to read

• Antigravity (Google)
• Terminal-Bench
• Agent evaluation
• WebArena
• DeepSWE
• AutomationBench

Topics:AI AgentsAI BenchmarksModel Evaluation

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10SemiAnalysis. GDPx version 1.0 leaderboard. Default scoring description, selected model scores, displayed run costs and eight-domain exclusion filter, accessed November 4, 2026.
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Kan, Max; Megalaa, Andrew; Patel, Dylan. GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains. Launch article and methodology, November 2, 2026.
3. ^1 ^2 ^3 ^4 ^5 ^6SemiAnalysis. GDPx domains. Domain definitions, partner assignments and task counts, accessed November 4, 2026.
4. ^1 ^2SemiAnalysis. GDPx rollouts. Public transcript selection and release scope, accessed November 4, 2026.
5. ^1 ^2OpenAI. Measuring the performance of our models on real-world tasks. GDPval release, task design and limitations, September 25, 2025.