Gemini 4 Carbon Performance Analysis: From Vulnerability Detection to Autonomous Coding

Google announced Gemini 4 Carbon on October 28, 2026, emphasizing stronger cybersecurity analysis and autonomous coding across complex, multi-step engineering tasks. It is designed to find software vulnerabilities, evaluate whether flaws are exploitable, and help produce fixes. As of November 1, 2026, however, it is not broadly available to the public.

🔎 Key takeaways

• ✅ Vulnerability discovery: 88.2% on SEC-bench Pro

• ✅ Exploitability evaluation: 54.8% on ExploitGym

• ✅ Vulnerability patching: 71.0% on CWE-bench

• ✅ Autonomous coding: built for extended, multi-step engineering workflows

• ✅ Availability: initially restricted to selected cyber defenders

🛡️ How much better is vulnerability detection?

Carbon is designed to support more of the security workflow, from identifying suspicious code to validating risks and proposing working fixes. This is more ambitious than simply highlighting a possible bug.

In Google’s published evaluations, Carbon scored 88.2% on SEC-bench Pro, versus 83.4% for Gemini 4 Argon—a 4.8-percentage-point difference. It also scored 54.8% on ExploitGym and 71.0% on CWE-bench, which assess different aspects of risk validation and vulnerability remediation.

These are benchmark scores under specified test conditions. They do not establish that the model can find or repair every vulnerability in a production environment.

💻 Can autonomous coding work without constant instructions?

Autonomous coding means an AI agent can break an engineering goal into steps, inspect files, modify code, run tests, and use the results to decide what to try next. The model is intended to sustain that workflow rather than stopping after a single code suggestion.

Consider a hypothetical example: a large application repeatedly fails in one module. An AI coding agent could inspect related files, identify a likely cause, propose a patch, and test it before reporting what happened.

Google says internal teams observed Carbon working through difficult engineering problems for days, including subtle bugs in large codebases. On Terminal-Bench 4.0, Carbon scored 65.2%, compared with Argon’s 57.1%.

That does not eliminate the need for developers. Production deployment, access permissions, and critical changes should still have human review and appropriate testing.

📊 Gemini 4 Carbon vs. Argon: benchmark results

Carbon leads Argon on each of these published measures. Real-world performance may differ depending on the codebase, available tools, permissions, and testing conditions.

🔐 Is it safe to let AI change code?

Google reports a 0.0% attack success rate for Carbon on a particular indirect prompt-injection benchmark. Its internal tests also found roughly half as many unsupported claims as Argon.

A zero score for attackers in one evaluation does not prove immunity to all attacks. Organizations should continue to review changes, restrict permissions, and run independent checks before merging or deploying generated code.

🚀 When can developers and consumers use Carbon?

As of November 1, 2026, access starts with a selected group of cybersecurity defenders. Google has said broader availability for developers, businesses, and consumers is planned, but has not announced a firm general-release date.

The announced introductory API price is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the introductory period. Published pricing should not be confused with immediate public access.

✅ What should you do next?

The important shift is not only faster code generation, but greater ability to carry a security or engineering task from investigation to validation and remediation. Identify one recurring security review or bug-fixing workflow in your own development environment. That gives you a practical baseline against which to assess Carbon when access becomes available.