Google Unveils Gemini 4 Carbon: Mixed Impressions, Unexpected Cybersecurity Prowess, and Pricing Puzzles
Google’s highly anticipated Gemini 4 Carbon model has been released, initially positioned as a “frontier performance” leader. While official benchmarks, such as Artificial Analysis, awarded it a score of 60, surpassing Opus 55, early real-world testing has revealed a more complex picture. For practical code analysis and review, the model’s performance has been underwhelming, exhibiting a notable hallucination rate. Although this rate is a 50% reduction from its predecessor, Argan (decreasing from 36% to 18%), it still falls short compared to other leading models. Furthermore, a significant point of concern is Gemini 4 Carbon’s speed, consistently ranking as the slowest among tested models and rendering its use for intensive tasks a “miserable” experience, despite claims of high tokens per second.
Economically, Gemini 4 Carbon introduces a dynamic pricing model, with an introductory rate set to double after December 1st, potentially making it more expensive than Opus for often inferior performance. Compounding this, the model’s caching mechanism is reportedly unreliable, leading to frequent cache misses and up to a 50% increase in requests, thereby escalating the actual cost per task. However, a surprising strength emerged in specialized domains, particularly cybersecurity and incident response. The model demonstrated exceptional capability in identifying and proposing practical solutions for complex security vulnerabilities, outperforming other advanced models that either declined assistance or offered impractical fixes. This aptitude aligns with Google’s emphasis on “Frontier cyber security capabilities.” While not positioned as a default for general-purpose coding where Opus remains the preferred choice, Gemini 4 Carbon represents a “significant jump in the right direction” for the Gemini line, showcasing real progress in niche applications and benefiting from the DeepMind team’s responsiveness to developer feedback.