What matters in AI.

Subscribe

Gemini 4 Carbon gets 10.9% on GDPx, near Claude Opus 5.5 at 12%

In tests of SemiAnalysis, the cost of a task is $16.29 with Carbon and $51.41 with the model of Anthropic.

SemiAnalysis made GDPx, a new test with 206 hard tasks in 11 fields, with 14 specialist partners. On Monday its researchers wrote that Gemini 4 Carbon got 10.9%, near the 12% of Claude Opus 5.5 of Anthropic. The model GPT-6 Astra of OpenAI got less than Carbon, and Carbon had a much lower cost for each task.

How the test works

GDPx has tasks in fields such as semiconductor design, finance and science. SemiAnalysis made it because other tests are too easy for top models. The top model got only 12%, and a new model can get more.

The numbers

  • Carbon: 10.9% at $16.29 for each task
  • Opus 5.5: 12% at $51.41 for each task

Carbon had the top result in software engineering and biology. The researchers tell that its biology result can come from safety filters that limit less.

What is not known

Carbon is not available to many users, and the first price can change. No GDPx score is given for GPT-6 Astra. Traders on Polymarket give Google 57% to have the top model at the end of 2026, but that market uses the Arena leaderboard.

Insight: Each lab leads on the test it selects

Google tells that Carbon is first on the Artificial Analysis index at 59. At this time it gives Carbon only to a small group of cyber defenders. Price and access, more than rank, seem to set who gets the customers, because each lab has the top result on some test. Google expects capital spend of $195 billion to $205 billion this year, which reads as a push to get customers.

Sources

  1. Google's Latest AI Model Nearly Matches Anthropic at One-Third the Cost: ReportYahoo Finance
AI MATTER · NEWS · AI MATTER · NEWS ·4 NOV2026

Posted

Tags

More in Research

All Research news