Google has a new frontier AI model, and this time the pitch is not just about writing better code. It is about finding the flaws in code before attackers do, and then fixing them. The company unveiled Gemini 4 Carbon on Wednesday, only two weeks after it introduced Gemini 4 Argon, and it is already calling Carbon its strongest frontier model yet.
The announcement came from Google chief executive Sundar Pichai and from Google DeepMind, which described Carbon as a model built for long, complex engineering work. It can stay with a technical problem for days, DeepMind said, tracing the root cause of a bug and working through fixes across large codebases and low-level systems.
Carbon will not reach everyone at once. Google said a select group of cybersecurity defenders will get access first this week, with access then widening to trusted partners through its Fairwind Program. Developers, businesses and consumers will follow later, after Google has gathered feedback and tested its safeguards.
A clear lead on security benchmarks
The headline numbers are about cybersecurity, and they are strong. On SEC-bench Pro, a benchmark for finding vulnerabilities in real-world software, Carbon scored 88.2 per cent. On ExploitGym, which tests whether a vulnerability can actually be exploited, it scored 54.8 per cent. And on CWE-bench, which measures how well a model develops fixes, it scored 71 per cent. Google said Carbon holds a firm lead across all three.
On the wider Artificial Analysis Intelligence Index, updated to version 5.0, Carbon scored 59 and ranked first, ahead of Anthropic’s Claude Opus 5.5 at 58 and OpenAI’s GPT-6 Astra at 57. That ranking covers coding and agentic tasks as well as cybersecurity, and it puts Carbon at the top of a very crowded field, at least for now.
Google also pointed to its own internal testing. Carbon generates secure code at more than twice the success rate of Argon, the company said. Those are Google’s own figures, so they should be read as the company’s claims rather than independent results.
Why cybersecurity comes first
There is a practical reason for the security-first rollout. AI models that can find and exploit vulnerabilities are powerful tools in the wrong hands, and Google knows it. By starting with defenders, the company is trying to put the capability in the hands of the people who patch systems before it spreads more widely.
DeepMind said Carbon has been built to behave differently when a task is ambiguous or high pressure. Instead of guessing, it seeks human input. In internal evaluations, it resisted every evaluated attempt at indirect prompt injection, where hidden instructions try to steer a model off course.
What it means for developers and businesses
For developers, Carbon points to a near future in which security checks are woven into everyday coding rather than bolted on at the end. A model that can scan a large codebase, find a real flaw, prove it can be exploited, and then write a fix changes how teams think about review. It does not replace human review, but it could catch the problems that slip through when deadlines are tight.
For businesses, the early message is narrower. Most companies will not touch Carbon directly for some time. When access does open up, introductory pricing is reported at 2 dollars per million input tokens and 10 dollars per million output tokens, matching Argon’s introductory pricing.
A fast-moving race
Carbon arrives in the middle of an unusually fast stretch for frontier models. Argon landed two weeks ago, and OpenAI has released GPT-6.1 Terra. The gaps at the top are now small, often a single point, and they are unlikely to stay still for long.
What sets Carbon apart, for now, is the combination: top scores on vulnerability discovery, exploitation testing and patching, plus a rollout plan that starts with the defenders. Whether that lead survives independent testing and wider use is the question the next few months will answer.
Frequently asked questions
What is Gemini 4 Carbon?
Google’s latest frontier AI model, unveiled on October 29, 2026, for complex coding and engineering tasks, with a focus on finding software vulnerabilities, testing whether they can be exploited, and developing fixes.
How does Gemini 4 Carbon compare with rival models?
Google said it ranks first on the Artificial Analysis Intelligence Index version 5.0 at 59, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57, and leads three cybersecurity benchmarks at 88.2 per cent, 54.8 per cent and 71 per cent.
When will Gemini 4 Carbon be available to everyone?
Not yet. Defenders get access first this week, then trusted partners through the Fairwind Program. Developers, businesses and consumers follow later, after feedback and safeguard testing.
How is Gemini 4 Carbon different from Gemini 4 Argon?
Argon launched two weeks earlier. Google said Carbon generates secure code at more than twice Argon’s success rate on internal benchmarks and cuts unsupported claims by about half.
How much will Gemini 4 Carbon cost?
Introductory pricing is reported at 2 dollars per million input tokens and 10 dollars per million output tokens, matching Argon’s introductory pricing.
Google’s next frontier model is here, and its first job is defence. The wider test will be whether Carbon earns the trust of the developers and defenders who use it every day.
