Hermes Vector

AI & SI

Google announces Gemini 4 Carbon frontier AI model

Google has introduced Gemini 4 Carbon, a new frontier AI model focused on cybersecurity, coding, and agentic tasks, currently undergoing voluntary testing.

By Hermes-Vector AI Desk

Checked against 2 sourcesArs TechnicaGoogle
Illustration for: Google announces Gemini 4 Carbon frontier AI model

In brief

  • Google announced Gemini 4 Carbon, a frontier AI model focused on cybersecurity and agentic tasks.
  • The model reportedly shows a significant leap in vulnerability detection, validation, and remediation.
  • Google is using a voluntary testing framework with the US government before a wide release.
  • Introductory API pricing is set at $2 per million input tokens and $10 per million output tokens.

Google announces Gemini 4 Carbon frontier AI model

Google has announced Gemini 4 Carbon, its latest frontier AI model. The company describes Carbon as its "strongest model with state-of-the-art cybersecurity and long horizon capabilities."

Performance and Benchmarks

Google states that Carbon shows significant gains in coding, agentic tasks, and scientific analysis. According to Google, the model is twice as likely to generate code free of vulnerabilities compared to its predecessor, Gemini 4 Argon.

In specific benchmarks, Google reports the following for Carbon:

  • SEC-bench Pro: 88.2% for finding vulnerabilities.
  • ExploitGym: 54.8% for exploitation.
  • CWE-bench: 71.0% for patching.
  • WebDev Arena: An increase of 89 points, ranking at number four on the leaderboard.

Google also claims Carbon ranks #1 on the v5.0 of the Artificial Analysis Intelligence Index at 59.

Cybersecurity and Safety

The Chrome team claims that Carbon is 10 times better at generating "high-quality" patches compared to non-Google models. In internal testing, Google reports that Carbon's high-severity misalignment alerts dropped by almost 80%.

Regarding safety, Google reports that Carbon achieved a 0% score in the Grey Swan evaluation, resisting every attempted prompt-injection attack. This is an improvement over Argon, which had a 0.7% attack success rate, Claude Opus 5.5 at 1%, and GPT-6 Astra at 8.5%.

Rollout and Pricing

Google is taking a "deliberative approach" to the release of Carbon. The company is participating in a voluntary testing framework established by the US government and plans to roll the model out through its Fairwinds cyber defense program. Access will initially be provided to a subset of cyber defenders and trusted partners.

For developers and AI Ultra subscribers, Google will offer an introductory API rate of $2 per million input tokens and $10 per million output tokens. Cached input tokens will be priced at 95% off the input token price. Google has not specified when the introductory pricing period will end.

Sources

This article was drafted with AI assistance and checked against the sources above. Company claims are reported as claims. Cover image is AI-generated.

Questions readers ask

What are the specific benchmark scores for the Gemini 4 Carbon model?
Google reports the following benchmarks for Carbon: 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, 71.0% on CWE-bench for patching, and an 89-point increase on the WebDev Arena leaderboard. Additionally, the model ranks #1 on the v5.0 of the Artificial Analysis Intelligence Index at 59.
How does Gemini 4 Carbon compare to its predecessor and other models in cybersecurity?
The model is twice as likely to generate code free of vulnerabilities compared to Gemini 4 Argon and is 10 times better at generating high-quality patches compared to non-Google models. In internal testing, high-severity misalignment alerts dropped by almost 80%, and it achieved a 0% score in the Grey Swan evaluation, resisting every attempted prompt-injection attack.
Who will have initial access to the Gemini 4 Carbon model?
Access will initially be provided to a subset of cyber defenders and trusted partners as part of Google's Fairwinds cyber defense program.

The day's checked news, by email

One short email a day with the top stories. Free, and you can unsubscribe in one click.

Or follow for the latest news:

Advertisement

How do you feel about this story?

Comments 0

No comments yet. Start the discussion.

More AI stories