Gemini 4 Carbon Takes Aim at Harder Engineering Work

Fixing a difficult software bug can take much longer than writing the first plausible patch. Someone has to trace the failure, test competing explanations, and check that the repair preserves everything else. Gemini 4 Carbon is Google's latest attempt to make AI more useful across that entire process. Announced on October 28, 2026, it arrives with stronger engineering and cybersecurity results, but access initially goes to selected cyber defenders rather than the wider developer community.

For teams considering their next model, Carbon raises two immediate questions: how much stronger is it, and when can they actually evaluate it?

What kind of model is Gemini 4 Carbon?

Carbon builds on Gemini 4 Argon. Google reports improvements in agentic coding, computer use, multimodal understanding, and long-context reasoning. These capabilities can work together: an engineering assistant might need to interpret a diagram before investigating a failure in the software it describes.

Artificial Analysis's Carbon model profile lists a one-million-token context window, support for text, image, speech, and video input, and text output. The context window limits the combined input and output a model can handle in a single request. For an engineering assistant, that could mean considering source files alongside diagrams, screenshots, and a lengthy discussion of the problem.

Those input capabilities do not make Carbon an image or video generator. Its listed output is text. Likewise, a large context window does not establish how reliably it will retrieve a small but crucial detail from a repository. That needs a task-specific test.

Consider a bug that spans a parser and the code that calls it. An agent would need to trace the failure across both, propose a fix, and run regression tests. The useful outcome would be a clear explanation of the root cause and a patch that survives review. A longer run is worthwhile only if it produces a better fix.

What the benchmarks actually show

The strongest broad signal comes from an independent evaluator. On October 29, Carbon at the High reasoning setting leads the Artificial Analysis Intelligence Index v5.0 with a score of 59. The index combines nine evaluations covering knowledge work, coding, science, and general capabilities. Its four categories each contribute a quarter of the overall score.

That result supports a broad capability claim. It does not mean Carbon will win every coding task, and comparisons need to retain both the index version and reasoning setting. A score on an older index measures a different collection of tests.

For defensive security, Collinear's CWE-bench v1 provides a more concrete comparison. Its leaderboard reports the following programmatic pass@1 results:

The evaluation uses 120 audit-and-patch tasks, high reasoning, and a one-hour limit per run. Each model gets four runs per task; pass@1 estimates success on a single attempt. Success requires blocking the exploit while keeping existing tests passing. Carbon's gain over Argon is three percentage points. The harness is the software that lets an agent use tools and work on code, so these scores reflect the model and its surrounding setup.

Google says the cybersecurity results highlighted in its announcement were evaluated without guardrails. Its evaluation methodology also explains that the 88.2% SEC-bench Pro result comes from its own tests on the 183-task May 2026 JavaScript engine split. Google used Antigravity at the highest thinking setting and averaged pass@1 over three independent attempts per task. That result should not be compared directly with scores from a different task set.

For a security team, CWE-bench is especially relevant because its grader checks both the repair and the existing functionality. Google's SEC-bench Pro result addresses vulnerability discovery under a different setup. Each helps answer a different question about what an agent can do.

Access and pricing come with different timelines

As of October 29, Google's announced rollout begins with selected cyber defenders, followed by additional trusted partners through the Fairwind Program. The launch announcement gives no firm date for broader access for developers, businesses, and consumers.

Google lists introductory pricing of $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input. That discount implies $0.10 per million cached input tokens. After the introductory period, the announced rates rise to $4 for input and $20 for output. The announcement does not specify when Carbon's introductory period ends.

For a simple budgeting example, one million uncached input tokens plus 100,000 output tokens would cost $3 at the introductory rates and $6 afterward. This calculation covers only those token charges. A long-running agent's total operating cost also depends on how much work it performs and the infrastructure around it.

The Fairwind Program prioritizes defenders such as governments and critical infrastructure operators. Google vets applicants, and partners cannot redistribute model access. An application must be approved before an organization can gain access through the program.

Where Carbon could earn its place

For an engineering team, a useful first trial would be a private, previously resolved bug with the original fix withheld. Give the agent the repository as it stood before the repair and clear acceptance criteria, then review its patch and test results. Measure the cost of completing the task alongside correctness.

That trial would answer a question the headline scores cannot: can Carbon carry a difficult investigation through to a correct fix in your codebase? The published results make it a promising candidate for demanding engineering and defensive security work. A decision to switch models should follow evidence from those tasks.