Google Unveils Gemini 4 Carbon: The New AI Powerhouse for Cybersecurity
Google launches Gemini 4 Carbon, a frontier AI model topping benchmarks and specializing in software vulnerability detection and fixing.
Google has officially pulled back the curtain on Google Gemini 4 Carbon, its latest frontier artificial intelligence model designed to push the boundaries of technical engineering and cybersecurity. In a move that underscores the company's rapid release cycle, this new model arrives just two weeks after the launch of Gemini 4 Argon, positioning itself as a specialized tool for tackling complex, multi-step software challenges.
Topping the Charts: Unmatched Benchmark Performance
The Gemini 4 Carbon model has already made waves in the AI community by securing the top spot on the Artificial Analysis Intelligence Index with an impressive score of 59. However, its true strength lies in its ability to handle deep technical labor. According to Google DeepMind, Carbon is capable of working through intricate technical problems for days at a time, allowing it to identify root causes of bugs and implement fixes across massive codebases and low-level systems.
The model's proficiency in cybersecurity is backed by striking data across several key benchmarks:
- SEC-bench Pro: Scored 88.2 per cent in identifying software vulnerabilities.
- CWE-bench: Achieved 71 per cent in developing effective fixes for those vulnerabilities.
- ExploitGym: Recorded a 54.8 per cent success rate in testing whether vulnerabilities could actually be exploited.
A Leap Beyond Argon
Google CEO Sundar Pichai, announcing the release on X, highlighted that Carbon represents the current pace at which the company is pushing the AI frontier. When compared to its predecessor, Gemini 4 Argon, Carbon shows a dramatic improvement in security and reliability. Internal benchmarks reveal that Carbon is more than twice as successful as Argon in generating secure code.
The impact is even more pronounced in real-world applications. Google's Chrome team reported that when generating high-quality patches, Gemini 4 Carbon performed over 10 times better than a competing frontier model from another AI laboratory.
Safety First: Reducing Hallucinations and Cheating
As AI models become more capable, the risk of misalignment and "hallucinations" increases. Google has integrated strengthened safeguards into Carbon to mitigate these risks. Internal testing shows that Carbon has reduced unsupported claims by approximately half compared to Argon.
Furthermore, the company reported a significant drop in high-severity misalignment alerts within internal coding environments, which fell by nearly 80 per cent. In terms of integrity, Carbon's rate of successful cheating on Terminal-Bench 4.0 dropped to zero, with cheating attempts plummeting to just 0.4 per cent. Additionally, it recorded a zero per cent attack success rate on the Gray Swan indirect prompt-injection benchmark during evaluations.
Availability and Pricing
Google is adopting a cautious, phased rollout for Gemini 4 Carbon. The rollout plan is as follows:
- Initial Phase: Available this week to a select group of cybersecurity defenders.
- Expansion Phase: Access will be extended to trusted partners via the Fairwind Program.
- General Release: Once feedback is gathered and safeguards are further refined, the model will be opened to developers, businesses, and general consumers.
In terms of cost, Google is keeping the pricing consistent with Gemini 4 Argon. The model will be priced at USD 2 per million input tokens and USD 10 per million output tokens. To encourage efficiency, cached input tokens will be available at a significant 95 per cent discount.