Google introduced Gemini 4 Carbon, a model designed to detect software vulnerabilities, assess exploitability and generate code fixes automatically.

According to Google, Carbon ranked first in the Artificial Analysis Intelligence Index v5.0 during benchmark comparisons across vulnerability analysis tasks.

Model capabilities

The new model is intended to identify weak points in source code, evaluate potential attack vectors and propose concrete remediation patches.

It can generate fixes and suggest defensive changes while aiming to preserve existing functionality, according to Google’s technical description.

Testing results

In internal evaluations, Google reports several improvements of Carbon compared with Gemini 4 Argon across safety and reliability metrics.

• Secure coding: Carbon produced secure code more than twice as often as Gemini 4 Argon in the company's internal tests.
• Assertions: The model was about half as likely to output ungrounded assertions, reducing the frequency of unsupported claims roughly twofold.
• Robustness: In the Gray Swan test, no attempt to induce execution of malicious instructions embedded in external data succeeded against Carbon.

Availability and pricing

Mass access will not be immediate; Google will initially provide Carbon to cybersecurity specialists and trusted partners only.

Wider availability is planned after additional safety reviews and further testing, according to the company statement.

The pricing matches Gemini 4 Argon at $2 per million input tokens and $10 per million output tokens.

Index context

The Artificial Analysis Intelligence Index v5.0 evaluates models on tasks such as vulnerability detection, exploitability assessment and remediation proposal generation.

Google cites the index placement to illustrate Carbon’s comparative performance on automated code-analysis benchmarks used by industry evaluators.