The Extended Brief

GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains

Brief by The AI News AI newsroom · Nov 1, 2026, 9:02 PM EST edition

Original reporting by SemiAnalysis (Dylan Patel) — Max Kan · published Nov 1, 2026, 8:51 PM EST

A new benchmark scores the best frontier model at just 12%, arguing progress should be measured only in the high-ROI domains where AI revenue will concentrate first.

Key points

• Even the best frontier model scores just 12% on GDPx, a new benchmark for general agentic capability. source ↗
• GDPx covers 11 high-ROI domains—including chip design, cybersecurity, quant trading, and RSI—while skipping government, real estate, and insurance. source ↗
• Its authors argue GDP-share or O*NET-based benchmarks misjudge where the first trillion dollars of AI revenue will concentrate. source ↗
• GDPx says safety classifiers dragged down Opus 5.5 and GPT-6 Astra scores in cyber, biology, and alignment. source ↗
• The authors say OpenAI and Anthropic dominate, with Google's Gemini 4 Carbon the only model in the same ballpark. source ↗

The data

Numbers from the original article, machine-verified against its text

Practical applications

• If you evaluate models for cyber, biology, or alignment workloads, test safety-classifier refusal rates directly, since GDPx attributes several top models' low domain scores to classifier behavior rather than capability.
• Compare candidate models on GDPx's per-domain scores instead of the aggregate when your workload falls in one covered domain, because the authors report meaningful spikiness across domains.
• If your team tracks frontier progress with GDP-weighted or occupation-based benchmarks, run your shortlisted models against GDPx to check whether the high-ROI-domain ranking changes your picks.

Context

GDPx is an evaluation benchmark for agentic AI capability—multi-step task performance rather than single-question answering. Its authors break from benchmarks that weight tasks by share of GDP or the O*NET occupational taxonomy, betting that early AI revenue concentrates in a few high-ROI sectors. Safety classifiers, which restrict outputs in sensitive areas like cyber and biology, can depress measured scores even when underlying capability is high, according to the authors.

What to watch

• Gemini 4 Carbon's general availability will test whether its benchmark standing matches broader user impressions, as the authors themselves flag.
• Watch whether OpenAI or Anthropic dispute the classifier-sandbagging claim or publish results excluding cyber, biology, and alignment.

Related briefs

• Open-weight model GLM 5.3 built full Firefox exploit chains with zero human input in under 48 hours
• Large language models detect methodological errors that human peer review misses in clinical trials
• Anthropic's Unreleased AI Model Chains Flaws Into Working Chrome, Safari, Windows and Linux Exploits
• A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis

This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.