Google Opens Gemini 4 Carbon to Ultra Subscribers and Developers

Google has just launched Gemini 4 Carbon, the company’s most powerful frontier model to date. It arrives with advanced long-horizon agentic workflows, elite coding skills, and ground-breaking cybersecurity capabilities designed to find and fix complex software vulnerabilities.

The release follows Google’s earlier introduction of Gemini 4 Argon. While Argon targeted general complex problem-solving, Carbon pushes the performance envelope further, climbing to the top tier of industry benchmarks, including the Artificial Analysis Intelligence Index.

Developers and enterprise clients can access the model through Google’s API, priced at $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95%.

What sets Gemini 4 Carbon apart from previous iterations is its specialized focus on software security and defensive engineering. According to Google DeepMind, the model excels across major external security benchmarks evaluated without guardrails.

Carbon achieved a remarkable 88.2% on SEC-bench Pro, successfully uncovering hidden vulnerabilities buried deep within real-world software codebases. Beyond merely spotting flaws, the model secured a top score of 54.8% on the ExploitGym benchmark to validate risk through safe exploitation.

In addition to offensive and defensive code analysis, Carbon has shown near-perfect resilience against adversarial threats. Building on the strong baseline of its predecessor, Carbon achieved a 0.0% attack success rate on Gray Swan’s Indirect Prompt Injection Benchmark, successfully resisting every evaluated prompt injection attempt.

Beyond security applications, Gemini 4 Carbon delivers substantial upgrades for general software engineering and web development. Google noted that the model yields superior code quality and artifact generation, jumping 89 Elo points to land near the top of the WebDev Arena leaderboard.