Google Gemini 4 Carbon performs well in new benchmark tests, offering better value than Claude models

·by Henderson·Gemini
Google Gemini 4 Carbon performs well in new benchmark tests, offering better value than Claude models

Google's newly launched cutting-edge model, Gemini 4 Carbon, brings some exciting improvements. Although it did not top the new benchmark tests, its value for money has significantly increased. AI research and analysis firm SemiAnalysis recently released its new frontier-level benchmark GDPx, which includes data on Google's Gemini 4 Carbon model. The ranking covers major frontier models from OpenAI, Anthropic, and Meta, detailing the scores and costs of each model.

The benchmark test covered 206 tasks across various fields, including finance, materials science, cybersecurity, software engineering, law, biology, and more.

The results show that Anthropic's Claude Opus 5.5 easily took the top spot, with Google's Gemini 4 Carbon following closely behind. The gap between the two is relatively small, but cost is the most significant consideration here. The cost per task for Claude is approximately US$51 (about HK$398), while Gemini is just slightly above US$16 (about HK$125). Gemini also outperformed the next model on the ranking, GPT-6 Astra, which scored lower and had a cost per task of about US$54 (about HK$421).

The report notes that removing cybersecurity, biology, and adjustment fields would change the results. The security measures of Claude and GPT models allowed GPT-6 Astra to rank higher on the list, while GPT 6.1 Sol and Claude Sonnet 5.5 ranked above Gemini 4 Carbon. SemiAnalysis stated that this aligns more closely with their subjective impressions of each model. As for Google's existing frontier model, Gemini 4 Argon, it ranked eighth on both lists, surpassing Meta's Muse Wonder but falling below Wonder Full. Argon had a cost slightly lower than Carbon in this benchmark test.

Full. Argon had a cost slightly lower than Carbon in this benchmark test.

Google Gemini 4 Carbon excels in benchmark tests

When focusing on cybersecurity, both of Google's models outperformed those from OpenAI and Anthropic, but lagged behind Meta's Muse model, which showed significantly better performance than the others in this area. Google announced Gemini 4 Carbon on October 28, claiming significant improvements, although a specific public release date has yet to be confirmed, with expectations for a rollout to most users soon.

Google's own benchmark tests indicate that Gemini outperformed GPT-6 Astra and Opus 5.5 in several areas.

ItemSpecification
Cost per task (Gemini 4 Carbon)about US$16 (about HK$124)
Cost per task (Claude)about US$51 (about HK$397)
Cost per task (GPT-6 Astra)about US$54 (about HK$421)

H
About the author
Henderson