At the end of last month, Google unveiled its latest frontier AI model, Gemini 4 Carbon, which features agentic capabilities and enhanced cybersecurity functions. Just two weeks prior, the Mountain View-based tech giant had released another agent-enabled AI model, Gemini 4 Argon. During the launch, Google presented various benchmark results to demonstrate the powerful performance of Gemini 4 Carbon. Now, SemiAnalysis has released results from its latest “GDPx” benchmark—a test designed to evaluate the agentic capabilities and cost-effectiveness of AI models—comparing the performance of several different models. On this benchmark platform, Google’s frontier AI model ranked second, while also standing out as the most cost-effective among the top five models.
Claude Opus 5.5 outperformed Gemini 4 Carbon in the SemiAnalysis GDPx benchmark.
SemiAnalysis shared the results of its latest frontier benchmark—which assesses an AI model’s ability to execute agentic tasks—on the platform X. The results compare AI models from organizations including OpenAI, Google, Anthropic, Meta, xAI, and other frontier AI labs. Anthropic’s Claude Opus 5.5 took the top spot with a score of 12 percentage points, narrowly edging out Google’s Gemini 4 Carbon, which scored 10.9 percentage points.
However, Gemini 4 Carbon stands out as one of the most cost-effective frontier AI models on the list. It costs $16.29 per agent task execution and generates 594,000 output tokens. Compared to Claude Opus 5.5—which incurs an average cost of $51.41 per task and produces 548,000 output tokens—Google’s frontier model offers significantly better cost-effectiveness.
OpenAI’s GPT-6 Astra ranks third with a score of 9.6, at a cost of $53.99 per task and an output of 275,000 tokens. It is followed by Meta’s Muse Wonder Full, which achieved a score of 9.0. Notably, Anthropic’s Claude Fable 5.1 leads the list with a score of 8.5, though it comes at a higher cost of $67.16 per task, with an output of 418,000 tokens. Gemini 4 Argon also appears on the list with a score of 7.3; it is even more cost-effective than Gemini 4 Carbon, with a per-task cost of just $12.79.
For reference, in October, Google released Gemini 4 Carbon, which features enhanced cybersecurity capabilities. Currently, however, this AI model is available only to a “select group of cybersecurity defenders,” with plans to gradually expand access to “more trusted partners” through Google’s Fairwind program. Google claims that, in internal testing, Gemini 4 Carbon achieved a success rate in generating secure code that was more than twice as high as that of Gemini 4 Argon.