GDPx aims to measure real uses of AI models

SemiAnalysis launches GDPx, a benchmark evaluating the agentic capabilities of models in high-value domains. The best model scores only 12% on this test, according to the November 2, 2026 publication.

SemiAnalysis has published GDPx, a benchmark designed to measure the agentic capabilities of AI models in domains considered most profitable, from coding to finance, including cyber, biology, law, and chip design. According to the November 2, 2026 publication, the best model achieves only a 12% score, or 1 out of 7. The message is clear: general rankings say little about a model’s real value when it comes to acting, deciding, and chaining complex tasks.

A benchmark focusing where money might concentrate

GDPx, standing for “General Agentic Benchmark,” was built with at least one specialized data partner per domain, then finalized by SemiAnalysis. The stated idea is simple: if the first major AI revenues are to come from high ROI sectors, then models must be tested on these specific grounds, not just on abstract scores or economic averages. The benchmark notably covers scientific research, hardware, cybersecurity, biology, finance, law, quantitative trading, chip design, software engineering, materials, and alignment.

This approach contrasts with evaluations that aggregate everything into a single indicator. SemiAnalysis believes that a benchmark based on GDP share or SOC/O*NET job classifications misses the mark, as it dilutes uses where AI can already generate strong value. In short, a model can be good everywhere without being useful where budgets are unlocked. Sometimes, the dashboard resembles a fridge temperature curve: reassuring but not very informative for production.

Scores mainly reveal gaps between uses

The figure highlighted by SemiAnalysis is stark: the best model scores only 12% on GDPx. The company also notes that some models, like Opus 5.5 and GPT-6 Astra, had their overall score penalized by their safety filters. Removing the cyber, biology, and alignment domains, where these classifiers can sometimes trigger, changes the ranking [to be verified], reminding us that a benchmark is never neutral in its construction.

Another key point: according to SemiAnalysis, Muse Wonder Full is currently the best available model for cyber, with a significant lead. The publication thus emphasizes an angle rarely highlighted in public rankings: security and safeguards can alter the interpretation of a result, especially when some use cases are sensitive or heavily regulated.

Why this benchmark matters for the AI market

GDPx is less about “who is the smartest” and more about “who can really get the job done.” For companies, this changes how they buy, test, and deploy models. A generalist tool may impress in demos, but if it fails when managing action sequences in a business environment, the cost quickly rises. Publishers will likely have to deal with finer, more sector-specific evaluations closer to the uses where revenues are at stake.

SemiAnalysis indicates it worked with specialized data partners in each domain while ensuring final QA and task iteration. This setup gives GDPx practical scope but also a limitation: like any recent benchmark, it must be confronted with other tests, teams, and future model versions. Still, the initiative points to a clear market trend: the battle is no longer just about model size but about their ability to produce useful results in high-value sectors.

Key points

  • GDPx targets high ROI domains.
  • The best model scores 12% according to SemiAnalysis.
  • Cyber, biology, and alignment can skew scores.
  • SemiAnalysis worked with specialized partners.
  • Muse Wonder Full is reportedly the best cyber model.
  • Publication dated November 2, 2026.

By the numbers

  • 12% — best model score on GDPx, according to SemiAnalysis, November 2, 2026.
  • 1 out of 7 — equivalent of the best score announced by SemiAnalysis.
  • 100,000 — subscribers of the publisher SemiAnalysis, according to the Substack page.
  • 12.6 K — views of the launch X post at the time of the provided capture.

To read