Gemini 4 Carbon: Google Claims Benchmark Gains as Access Remains Limited

Google has announced Gemini 4 Carbon, a new AI model it says makes major gains in coding, agentic tasks and cybersecurity. Google claims Carbon leads the Artificial Analysis Intelligence Index v5.0 and performs strongly on SEC-bench Pro, but access is currently restricted to selected cybersecurity defenders. That limited rollout means most developers and businesses cannot yet test the model against their own work.

The announcement puts Carbon forward as a step beyond Gemini 4 Argon, which Google released earlier. Its headline results are notable, but the available reporting does not independently establish the benchmark methodology or provide enough detail to verify every comparison. For now, Carbon's potential is clearer in Google's claims than in broad, public use.

What Google Says Gemini 4 Carbon Can Do

Google describes Carbon as a frontier model designed for demanding software engineering and agentic work - tasks where an AI system may need to plan, use tools and complete a sequence of actions. The company also highlights cybersecurity, including finding weaknesses in code and generating patches intended to fix them.

These capabilities matter because security work requires more than producing code that runs. A useful model must identify genuine vulnerabilities, avoid introducing new ones and generate fixes that preserve the software's intended behavior. Strong performance on specialized evaluations could make AI a more capable assistant for security teams, but benchmark results alone do not establish that a model is safe to rely on in production.

How Its Benchmark and Cybersecurity Claims Compare With Argon

Google says Gemini 4 Carbon takes the top position on the Artificial Analysis Intelligence Index v5.0, an updated evaluation that the company describes as more demanding. It also points to Carbon's results on SEC-bench Pro, a benchmark focused on security-related software tasks.

For cybersecurity, Google's reported figures include more than twice Argon's secure-code success rate and a greater-than-tenfold advantage over an unnamed competing model in quality-patch generation. Those numbers suggest a meaningful improvement if the tests are comparable and representative. However, the rival is not identified in the available reports, and they do not independently establish how the evaluations were run or how results were judged.

The comparison with Argon is useful as a company-reported measure of progress, not a guarantee of the same gains across everyday development. Benchmark scores can depend on task selection, test conditions and scoring rules. Independent evaluations and practical experience will help show whether Carbon's reported advantages hold up across different codebases and security workflows.

Who Can Access Carbon - and What Wider Availability Depends On

Google is initially providing Gemini 4 Carbon to selected cybersecurity defenders through its Fairwind Program. The staged approach gives a limited group access while Google gathers feedback and iterates on the model. The company says it intends to expand access, but has not provided a broad public release timeline in the available reports.

For developers and businesses, that means Carbon is not yet a model they can freely evaluate or adopt. Teams interested in its coding and security capabilities will need to wait for wider availability or access through the program. Until then, the benchmark claims offer an early indication of Google's direction, rather than evidence that Carbon is ready for general deployment.

Google's announcement signals a push toward more capable AI for software and cybersecurity. The next important test will be whether independent results and broader access confirm the gains. Follow Google's Gemini updates for news on availability and further detail about how Carbon's benchmark performance was measured.