SemiAnalysis launches GDPx, an agent benchmark where the best model scores 12%

Version 1.0 has 206 tasks across 11 domains, built with 14 specialised data partners[2]. The domains are chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering[1].

The selection is opinionated on purpose. Authors Max Kan, Andrew Megalaa and Dylan Patel argue that the first trillion dollars of AI revenue will be concentrated in sectors that can deploy frontier intelligence at high return, so a benchmark weighted by share of GDP or by occupation lists would track the wrong things[1]. They intentionally skipped industries like government, real estate and insurance[1].

The default time limit is eight hours per task[1]. Models run in their own harness where one exists, such as Claude Code for Opus 5.5, Codex for GPT-6 Astra and Antigravity for Gemini 4 Carbon, and in OpenCode otherwise[1]. The headline score is pass@1 averaged over three attempts, with all domains weighted equally and each partner weighted equally within its domain[1][2].

SemiAnalysis says it approached more than 30 RL environment companies and put every partner through the same automated quality checks for consistency, cheating and unfair verifiers, backed by manual review of hundreds of trajectories[1]. It claims the same process flags nearly half of Terminal Bench 4.0's tasks as broken[1].

Claude Opus 5.5 leads at 12.0% ($51.41 per task), followed by Gemini 4 Carbon at 10.9% ($16.29), GPT-6 Astra at 9.6% ($53.99) and Meta's Muse Wonder Full at 9.0%[2]. Further down, GPT 6.1 Sol scores 8.4% at $9.80 per task, while the budget options DeepSeek 4.1 Flash and GPT-6 Luna land at 5.4% and 5.1% for under a dollar a task[2]. Grok 4.7 sits at 3.9% and GLM 5.3 at 2.7%[2].

Most of GDPx is private. Each partner open-sourced the rollouts for a single task, and the rest are held back to stop labs hill-climbing on them[3]. That protects the benchmark but means outsiders can't rerun it. With 206 tasks split 11 ways and scores this low, small gaps near the top shouldn't be over-read, and because models run in their own harnesses, the results measure the agent product as much as the model.

Benchmarks like SWE-bench and Terminal Bench are getting crowded at the top. GDPx resets the scale: on hours-long work in chip design, finance or RL research, even the best model earns an overall score of just 12.0%. It also shows the cost spread clearly, with Gemini 4 Carbon getting within about a point of Opus 5.5 for roughly a third of the price per task[2].