Gemini 4 Carbon

Gemini 4 Carbon is Google DeepMind’s AI model for coding, reasoning and work that involves many steps. Google announced it on 28 October 2026, two weeks after Gemini 4 Argon.

Google says Carbon improves its ability to find and fix software vulnerabilities and sustain longer engineering tasks. The first rollout is for selected cyber defenders, followed by additional trusted partners through Fairwind.

Access starts with selected cyber defenders

Google plans to begin access for a subset of cyber defenders during the week of the announcement. It says it is gathering feedback, refining guardrails and participating in the US government’s voluntary process for testing models before release.

Wider access for developers, businesses and consumers is planned. The announcement provides an access plan rather than a public API model identifier or a date when everyone can use Carbon. Follow the current rollout terms when choosing a model for a project that needs to run today.

Introductory prices are $2 input and $10 output

Google announced these prices per million tokens:

Cached input costs 95% less than ordinary input at the introductory rate. Google’s announcement lists $4 input and $20 output after the introductory period. The end date is unspecified, so Carbon’s pricing should be checked against its own terms as access expands.

Its score of 59 uses the new Intelligence Index

Artificial Analysis lists Gemini 4 Carbon at the High setting with an Intelligence Index score of 59 on version 5.0. The index combines several tests into one score. Compare models on the same index version: older v4.3 observations elsewhere on this site retain their original test conditions and dates.

An aggregate score gives one view of performance. A useful model choice also depends on the task, access, cost and the tools needed to complete the work.

Google reports stronger cybersecurity results

Google reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching. These are vendor-reported evaluations run with guardrails off. Their test conditions differ from a deployed product with safeguards and access controls.

The company also describes internal engineering runs lasting days, including investigating bugs across large codebases and low-level systems. Those examples describe Google’s testing. Real projects still need review of the changes, their effects and the permissions given to an AI agent.