Yahoo Finance reported on November 4 that Gemini 4 Carbon scored 10.9% on SemiAnalysis’s GDPx benchmark, compared with 12% for Claude Opus 5.5. The reported gap is 1.1 percentage points, while the reported cost per GDPx task was lower for Carbon at Google’s introductory pricing. The results describe this benchmark, not a universal model ranking.
The reported GDPx scores and task costs
Yahoo Finance reported both the scores and task costs below. Carbon’s reported task cost is tied to Google’s introductory pricing.
| Model | Reported GDPx score | Reported cost per GDPx task |
| Gemini 4 Carbon | 10.9% | $16.29 |
| Claude Opus 5.5 | 12% | $51.41 |
The cost figures differ by $35.12 per task. Calculated from the reported amounts, Carbon’s task cost was about 68.3% lower in this comparison. These are reported costs for GDPx tasks, not general API rates; the Carbon figure uses introductory pricing.
How GDPx calculates its headline score
GDPx covers 11 domains. Yahoo Finance reported that the benchmark includes 206 tasks; SemiAnalysis says it worked with 14 specialist partners. Its headline metric is pass@1 across three attempts, with each domain weighted equally. Where a domain has multiple partners, their contributions are weighted equally too.
SemiAnalysis uses binary or partial grading depending on the domain. Tasks have a default time limit of eight hours, though some are shorter, and the models ran in their native harnesses when available. These rules make the headline percentage a composite benchmark score shaped by domain weights, grading and task conditions. They do not make it a success rate for ordinary, real-world requests. SemiAnalysis’s GDPx report describes the scoring and task setup.
GDPx leaders vary by domain
SemiAnalysis reports different leaders across fields: Claude Opus 5.5 leads in quantitative trading, hardware design and chip design; Gemini 4 Carbon leads in biology and software engineering; GPT-6 Astra leads in recursive self-improvement and materials science. The overall scores therefore do not mean one model leads every kind of work.
SemiAnalysis also says safety classifiers depress some Opus 5.5 and GPT-6 Astra scores, and that weaker classifiers may partly explain Carbon’s biology result.
