SemiAnalysis releases GDPx, an agentic benchmark of 206 tasks across 11 high-value domains where the top model, Claude Opus 5.5, scores just 12%

Background and the problem

OpenAI's GDPval, released in September 2025, is still the general "knowledge work" benchmark most often reported on model release cards, but its tasks are one-shot and non-interactive: a few hundred words of instructions and an attachment, then the model writes a document. SemiAnalysis argues that this format predates agents and no longer separates today's frontier models, while many popular leaderboards are close to saturated. It also argues that the first trillion dollars of AI revenue will come from sectors that can deploy frontier intelligence at high ROI, so GDPx covers just 11 domains such as chip design, cyber, biology, finance, legal, quant trading and RSI, and deliberately skips government, real estate and insurance.

How it works

Each domain was built with at least one specialist RL environment company: Arithmetic built cyber tasks from proprietary zero-days in open-source software, Fleet built alignment tasks and Datacurve built software engineering tasks, with SemiAnalysis doing final QA and iteration. Tasks come with simulated work environments and a default 8-hour limit. Models run in their native harness where possible (Claude Code for Opus 5.5, Codex for GPT-6 Astra, Antigravity for Gemini 4 Carbon) and otherwise in OpenCode. Grading mixes binary and partial credit by domain; the headline is pass@1 over three attempts, with domains weighted equally.

Benchmarks and numbers

All figures are from the SemiAnalysis leaderboard: Claude Opus 5.5 12.0%, Gemini 4 Carbon 10.9%, GPT-6 Astra 9.6%, Muse Wonder Full 9.0%, Claude Fable 5.1 8.5%, GPT 6.1 Sol 8.4%, Claude Sonnet 5.5 8.2%, Gemini 4 Argon 7.3%, DeepSeek 4.1 Flash 1017 5.4%, GPT-6 Luna 5.1%, Grok 4.7 3.9% and GLM 5.3 2.7%. Nearly every model scores 0 on finance and legal. SemiAnalysis notes that Opus 5.5 and GPT-6 Astra trip safety classifiers in cyber, biology and alignment, which lowers their overall scores. The results have not yet been reproduced by a third party.

Getting started

Don't pick a model on the headline number alone: the site lets you filter by domain and by average cost per task and shows the Pareto frontier. SemiAnalysis's read is that GPT-6 Luna and DeepSeek compete at the low end, Opus wins the high end and OpenAI holds the middle, and that safety classifiers matter for cyber work. A top score of 12% is also a reminder that long, real-environment tasks still need human checkpoints. Data vendors and labs can contact [email protected] to join the next version.