Gemini 4 Carbon Frontier AI Model Falls Slightly Behind Claude Opus 5.5 in Agentic Capabilities, but Not in Cost Efficiency

Google launched its latest frontier AI model late last month, named Gemini 4 Carbon, with agentic and enhanced cybersecurity capabilities. The AI model was released just two weeks after the Mountain View-based tech giant released Gemini 4 Argon, another AI model with agentic capabilities. During the announcement, Google published multiple benchmark results to showcase the capabilities of the Gemini 4 Carbon. Now, SemiAnalysis has published results from its new GDPx benchmark that tests the agentic capabilities and cost efficiency of AI models, comparing results of multiple models. On the benchmarking platform, Google's frontier AI model came in second position, while being the most cost-efficient among the top five.

Claude Opus 5.5 Beats Gemini 4 Carbon on SemiAnalysis' GDPx Benchmark Test

In a post on X, SemiAnalysis published the results of its new frontier-level benchmark test, which tests how well AI models perform agentic tasks. The results compared the performance of AI models from OpenAI, Google, Anthropic, Meta, xAI, and other frontier AI labs. Anthropic's Claude Opus 5.5 topped the table with a score of 12 percentage points. The AI model outperformed Google's Gemini 4 Carbon frontier AI model by a small margin, which managed to score 10.9 percentage points.

However, the Gemini 4 Carbon was among the most cost-efficient frontier AI models on the list. It cost $16.29 per agentic task with 5,94,000 outputs tokens. Google's frontier model was significantly more cost-efficient compared to Claude Opus 5.5, which performed each at an average cost of $51.41 with 5,48,000 output token.

OpenAI's GPT-6 Astra took the third spot with a score of 9.6 percentage points, with a cost per task of $53.99 and 2,75,000 output tokens, followed by Meta's Muse Wonder Full with a score of 9.0 percentage points. Interestingly, Anthropic's Claude Fable 5.1 AI model scored 8.5 percentage points at the highest cost per task of $67.16 for 4,18,000 output tokens. Gemini 4 Argon also made it to the list with 7.3 percentage points. Argon was more cost-efficient than Gemini 4 Carbon, costing $12.79 per task.

For reference, Google launched the Gemini 4 Carbon with enhanced cybersecurity capabilities in October. However, the AI model is only available to a “subset of cyber defenders”, with plans to gradually offer phased access to “additional trusted partners” through Google's Fairwind Program. Google claims that Gemini 4 Carbon managed to achieve more than a 2x uplift in success rate for generating secure code over Gemini 4 Argon in internal testing