Google unveils Gemini 4 Carbon for security and coding work
Original title: Google says Gemini 4 Carbon beats its predecessors at finding and fixing security flaws
Access starts with selected cyber defenders via Fairwind at $2 per million input tokens before a later price rise.
Google says Gemini 4 Carbon, built on the Argon safety framework, is better than predecessors at finding flaws, judging exploitability and writing fixes. Google cites internal and external scores, including 88.2% on SEC-bench Pro, and says Carbon resisted every tested indirect prompt-injection attack on Gray Swan’s benchmark. Availability will begin with a subset of cyber defenders, then trusted partners, before a wider launch. Introductory pricing is $2 per million input tokens and $10 per million output tokens.
Unconfirmed: Performance, safety and prompt-injection results are Google’s reported evaluations, including tests run without guardrails.
- Google says secure-code success was more than twice Argon’s.
- Chrome team reported more than 10 times better high-quality fixes than another lab’s model.
- External scores: 88.2% SEC-bench Pro, 54.8% ExploitGym, 71.0% CWE-bench.
- Terminal-Bench 4.0 cheating success fell to 0%; attempt rate to 0.4%.
- Later pricing rises to $4 input and $20 output per million tokens.
- Vendors
- Organizations
- Google DeepMindWiz
- AI models
- Gemini 4 CarbonArgon
From the original article
2 min read at helpnetsecurity.com
Google’s Gemini 4 Carbon is a frontier AI model designed for software engineering and cybersecurity tasks, with safeguards built on those of its predecessor, Argon. SEC-bench Pro (Source: Google) Google says its teams have seen Carbon spend days investigating and fixing bugs across large codebases and low-level systems. The model was also trained to handle complex tasks and high-pressure situations.
By Anamarija Pogorelec · helpnetsecurity.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.