Gemini 4 Carbon strengthens Google in cybersecurity
Google introduces Gemini 4 Carbon, its new frontier model, with claimed gains in cybersecurity, coding, and long tasks. Deployment begins with a limited circle of cyber defenders before wider release.
Google unveiled Gemini 4 Carbon, its new frontier AI model, on October 28, 2026, with a very clear focus: cybersecurity. According to the company, this model succeeds Gemini 4 Argon launched two weeks earlier and also advances in coding, agent tasks, and multimodal analysis. For now, access remains limited to a small group of cyber defenders via the Fairwind Program, before broader distribution in Google products. The message is clear: the company wants to regain initiative in an area where raw performance matters, but caution weighs just as much.
A model designed to find flaws, then fix them
Google describes Gemini 4 Carbon as its most advanced model for spotting vulnerabilities and producing secure code. On SEC-bench Pro, a test measuring a model’s ability to discover real flaws, Carbon scores 88.2%; on ExploitGym, which evaluates how a model can turn a vulnerability into an exploit, it scores 54.8%. On Collinear CWE-bench, it achieves 71.0% for generating functional patches. The company also claims its internal tests show a success rate at least twice that of Argon for secure code generation. Incidentally, the model also performs well against prompt injection, a technique aimed at diverting its initial instructions.
This shift toward cyber is no trivial matter. Generative models can help find bugs but also accelerate attacks if safeguards are too weak. Google therefore emphasizes a set of protections: real-time abuse defense, reasoning chain monitoring, and resistance to manipulation. On Terminal-Bench 4.0, Carbon shows 0.0% success at cheating, with only 0.4% attempts. In internal coding work, high-severity misalignment alerts reportedly dropped by about 80%. The model also promises fewer unfounded claims, with half as many unsupported assertions as its predecessor. Not bad for a tool meant to say fewer nonsense things than it detects.
Why Google first reserves it for cyber defenders
The deployment of Gemini 4 Carbon follows a logic of progressive control. Google first opens it to a subset of cyber defense professionals this week, then will extend it to trusted partners via the Fairwind Program before integrating it into products for developers, businesses, and the general public. This sequence also reflects Google’s participation in the US voluntary pre-launch testing process. In short, the company prefers to verify safeguards before opening the floodgates.
The model is designed for long and complex workflows. Google says it has observed Carbon operating for days, persisting after dead ends, and completing multi-step engineering tasks in large codebases and low-level systems. On the WebDev Arena, its web performance reportedly gained 89 Elo points to reach 1,769, placing it fourth. On the Artificial Analysis Intelligence Index v5.0, it takes first place with a score of 59, ahead of Claude Opus 5.5 and GPT-6 Astra according to Google and Artificial Analysis.
Solid performance, but still tightly controlled
Comparison with Gemini 4 Argon shows rapid scaling. Carbon arrives just two weeks after its predecessor, illustrating Google’s fast pace in this segment. The announced gains cover reasoning, software development, and robustness in cyber contexts, with an ability to resist shortcuts and better collaborate with users. For businesses, the immediate stake is clear: a more reliable model can reduce time spent fixing, testing, or securing code. For attackers, the same progress is also a warning signal.
The question of access remains. No public release is planned at this stage, and Google has not provided a timeline for wider availability. The announced pricing matches Argon’s: $2 per million input tokens and $10 per million output tokens, with cached inputs billed 95% less than standard tokens. At this stage, Carbon is as much a branding product as a real-world test on the shifting frontier between power and control.
By the numbers
- 88.2% — SEC-bench Pro score of Gemini 4 Carbon, according to Google.
- 54.8% — score on ExploitGym, according to Google.
- 59 — score on the Artificial Analysis Intelligence Index v5.0.
- 1,769 — WebDev Arena score, after an 89 Elo point gain.
- $2 / $10 — announced prices per million input and output tokens.