Google has introduced its newest frontier AI model, Gemini 4 Carbon, claiming top performance against major industry rivals. According to The Verge, the model outperforms Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Astra on key industry benchmarks. However, everyday AI users looking to test Carbon in their browser or mobile app will find it completely out of reach for now.
Instead of launching the model inside its standard chatbot services, Google is restricting access to specialized corporate partners. Here is a closer look at what Gemini 4 Carbon does, why everyday users cannot access it, and what it costs for those who can.
High benchmark scores and specialized skills
Gemini 4 Carbon was built specifically for complex coding tasks, software engineering, and cybersecurity defense. According to The Verge, Carbon earned the top spot on the Artificial Analysis Intelligence Index (version 5.0), ranking ahead of competing models from OpenAI and Anthropic.
In security-specific evaluation tests, Carbon placed first on SEC-bench Pro, a benchmark that measures an AI system’s ability to identify real-world software bugs. It also led the ExploitGym benchmark, which evaluates how effectively an AI can transform security vulnerabilities into working exploits. Google reported that Carbon resists prompt injection attacks and avoids taking shortcuts when working through difficult problems.
The new release follows Argon, an earlier model Google launched earlier this month. Google says Carbon reduces unsupported claims by two times compared to Argon, providing greater accuracy and transparency. Internal teams at Google reported that Carbon could run for days to persist through dead ends to tackle complex, multi-step engineering tasks. These features make it well-suited for autonomous work, acting as an AI agent that works through complex workflows without requiring constant human guidance.
Why everyday users cannot try Carbon today
Despite Carbon’s high test scores, everyday consumers cannot try the model in the regular Gemini app today. Google has limited access exclusively to a curated list of cybersecurity defense companies. Moving forward, access will slowly expand to select partners participating in Google’s Fairwind Program.
Keeping a powerful cybersecurity model restricted is a safety choice. Because Carbon is specifically engineered to spot software vulnerabilities and construct security exploits, making it broadly available to the public could pose significant digital safety risks if bad actors misuse it. Restricting the tool to vetted defense organizations allows Google to test its capabilities in controlled settings.
What Gemini 4 Carbon costs and what comes next
For the select organizations granted access, Google has set an introductory price of $2 per million input tokens and $10 per million output tokens. That pricing matches the cost of Google’s earlier Argon model.
For everyone else, the standard Gemini experience remains available to anyone with a Google account. The free version offers basic text assistance along with image creation and editing. Additional capabilities, including video generation and higher usage allowances, are reserved for paid Google AI subscription plans.
Gemini 4 Carbon is not intended to replace everyday consumer tools in the immediate future. Instead, its focus on multi-day task management and advanced defense shows how tech companies are developing highly specialized AI models for enterprise security rather than general consumer chat.