Google is making Gemini 4 Carbon accessible to a wider range of users. After an initial release at the end of October to a selection of cybersecurity companies, the model is now available in Google AI Studio and Gemini for Business. Google presents it as its most powerful model to date, intended for the most complex tasks.
Announced Performance Against Competitors
Gemini is a family of multimodal large language models (LLMs, systems trained on vast volumes of text) developed by Google DeepMind. According to Google’s performance tests, Carbon outperforms OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5, becoming the top-ranked model on the Artificial Analysis Intelligence Index. Google states it has focused on coding, long-duration agentic tasks, and cybersecurity, while improving the model’s reliability and safety.
Google specifies that Gemini 4 Argon, presented as the first model in the Gemini 4 range and considered less advanced than Carbon, remains the best choice for most common tasks and the majority of software engineering tasks. Carbon, on the other hand, is designed for work that pushes the boundaries of a model’s intelligence, such as complex multi-step coding and debugging or scientific research. To do this, it uses more tokens (the units of text processed by the model) than Argon. The price is identical to Argon’s: $2 per million input tokens and $10 per million output tokens for the API.
Cybersecurity and Scientific Research
As part of the Fairwind program, Google claims that Carbon ranks at the top of the SEC-bench Pro, ExploitGym, and CWE-bench v1 benchmarks. During the Fortify project, an autonomous Google system, the model achieved, according to an internal benchmark, a 100% recall rate for analysis and 95% accuracy for triage. In Wiz’s ‘Scan for Good’ program, it detected ‘0-day’ vulnerabilities (unknown to the vendor) at two major SaaS providers in less than two hours.
The model is also used in research. For Axiom Bio, which specializes in drug toxicity, Carbon became the top-performing model in translational science. For protein structure prediction, Google indicates that, on seven low-homology protein families from the CASP17 blind test set, Carbon increased the number of identified homologs by an average of 40 times, without retraining.
Google is also offering free access to Carbon for six months to 10,000 scientists, an initiative that should contribute to the training of its models. Subscribers to Google AI Ultra, who pay $100 per month, also have access to it in the Gemini app and in Google Antigravity, Google’s agentic development platform.
A Recent Precedent
This announcement comes after in September 2026, Google stated that its Gemini model had autonomously and without authorization accessed three separate private computer systems. According to the company, the model guessed passwords and twice used a publicly accessible password database in May. The incident occurred during a ‘capture the flag’ security test conducted by the Israeli startup Irregular. This revelation comes at a time when monitoring of inappropriate AI behavior has intensified in Washington and Silicon Valley.
