Google’s Gemini 4 Carbon has ranked just behind Anthropic’s Claude Opus 5.5 in a new frontier-model benchmark, while offering a considerably lower estimated cost per task.
The GDPx benchmark, published by AI research and analysis firm SemiAnalysis, evaluates 206 tasks across areas including finance, materials science, cybersecurity, software engineering, legal work and biology. It compares major frontier models from Google, Anthropic, OpenAI and Meta by both performance and cost.
Claude Opus 5.5 topped the overall ranking, with Gemini 4 Carbon close behind. However, the cost difference was substantial: Claude was estimated at around $51 per task, while Gemini cost just over $16. OpenAI’s GPT-6 Astra ranked below both models and cost approximately $54 per task.
Removing the cybersecurity, biology and alignment categories changed the leaderboard. GPT-6 Astra moved into first place, followed by GPT 6.1 Sol and Claude Sonnet 5.5, with Gemini 4 Carbon ranking lower. SemiAnalysis said this adjusted result was more consistent with its subjective impressions of the models.
Google’s existing Gemini 4 Argon ranked eighth in both versions of the benchmark. It placed ahead of Meta’s Muse Wonder but behind Wonder Full, while its cost was only slightly lower than Carbon’s in this evaluation.
Google’s models performed better than those from OpenAI and Anthropic in cybersecurity-focused tasks, although Meta’s Muse models led that area by what SemiAnalysis described as a meaningful margin.
Google announced Gemini 4 Carbon on October 28, highlighting significant gains over previous models. Its public release date has not yet been confirmed, and the model is currently described as coming soon for most customers.
Sources: 9to5Google.