Introduction of GDPx Benchmark for Frontier Intelligence

A new benchmark designed to measure the capabilities of frontier artificial intelligence models, titled GDPx, has been released by SemiAnalysis. Positioned as a successor to the previous GDPval metric, GDPx shifts the focus from broad economic indicators, such as share of GDP or occupational taxonomies, toward high-value domains expected to drive the first trillion dollars in AI-related revenue.

According to documentation from SemiAnalysis, the benchmark specifically targets sectors where frontier intelligence is expected to yield high returns on investment (ROI). These domains include chip design, cybersecurity, biology, finance, legal services, quantitative trading, recursive self-improvement (RSI), hardware design, materials science, alignment, and software engineering. The developers have intentionally excluded industries such as government, real estate, and insurance from the scope of the benchmark.

Performance Metrics and Model Rankings

Initial data released via social media and SemiAnalysis suggests that current top-tier models are achieving scores of approximately 12% on the GDPx metric. While some observers on community forums like Reddit have speculated that these scores may saturate quickly as models improve, the current rankings provide a snapshot of the competitive landscape.

When excluding subjects that may trigger safety classifiers—such as cybersecurity, biology, and alignment—the performance rankings for leading models are as follows:

  1. Astra: 11.8%
  2. Opus 5.5: 11.7%
  3. Sol 6.1: 10.5%
  4. Sonnet 5.5: 9.7%
  5. Gemini 4 Carbon: 9.2%
  6. Fable 5.1: 8.6%
  7. Muse Wonder Full: 6.2%

The data indicates that the activation of safety classifiers can negatively impact model performance in certain domains, according to unverified reports circulating on social media.

Economic Implications and Value Capture

The emergence of GDPx coincides with broader industry discussions regarding the concentration of economic value within the AI sector. As model developers attempt to transition into platform providers, market analysts are questioning whether the primary financial gains will accrue to model creators or to the layers controlling compute and end-to-end application workflows.

Industry commentary suggests that durable profits in the AI sector may depend on deep integration into professional workflows, creating switching costs that extend beyond marginal improvements in benchmark scores. This debate follows recent market activity, including Anthropic’s IPO, which has intensified scrutiny over which participants in the AI stack will ultimately capture the most significant economic value.