Gemini 4 Carbon: Access, Pricing and Benchmarks

Google announced Gemini 4 Carbon on October 28, 2026 as a new frontier model focused on cybersecurity and long-running engineering work. The launch is significant, but the first practical question is availability: Carbon is not a generally available Gemini API model. Google is initially giving selected cyber defenders access, then plans to expand to trusted partners before a broader product rollout. The published prices and benchmark results should therefore be read as launch guidance, not as proof that any developer can use Carbon today.

Who can access Carbon now?

Google says access begins this week with a subset of cybersecurity defenders, with additional trusted partners phased in through its Fairwind Program. It has not published a universal release date, public API model identifier or guaranteed country-by-country availability. Fairwind is a vetted program aimed at defensive organizations, not an instant self-service sign-up for every developer. If your team needs a production model now, evaluate the separately released Gemini 4 Argon rather than designing an integration around unconfirmed Carbon access.

Carbon versus Argon: capability and rollout

The comparison supports a narrow conclusion: Carbon has stronger scores on these disclosed tests, but benchmark leadership does not establish lower production costs, better performance on every workload or immunity to agent failures. Google's earlier Argon rollout is independently documented and should not be confused with Carbon's staged access. A team building an automated coding assistant should compare task completion, human corrections, latency and security incidents on its own representative cases once Carbon is actually available.

Cybersecurity benchmark results and their limits

Google highlights three cybersecurity benchmarks for Carbon: SEC-bench Pro for vulnerability discovery, ExploitGym for evaluating exploitation capability, and CWE-bench for code patching. It reports 88.2%, 54.8% and 71.0%, respectively. Google also says its internal secure-code generation success rate was more than twice Argon's, and describes fewer severe misalignment alerts. These figures come from disclosed tests and vendor reporting, not a guarantee that Carbon will find or fix every issue in a real repository.

Announced pricing versus billable access

Google says Carbon will launch at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95%. After the introductory period, it lists $4 per million input and $20 per million output tokens. The announcement does not give a Carbon-specific end date for that introductory period. It also does not establish that Carbon is already billable in the public Gemini API. Check the live model catalog and billing documentation when your account receives access.

A practical evaluation plan for AI teams

1. Confirm your organization is eligible for an official Carbon access route; do not assume access from an Argon subscription or API key.
2. Choose 20–50 representative, authorized engineering or defensive tasks and define human-reviewed success criteria.
3. Compare an available baseline model on correctness, evidence quality, patch acceptance, latency and total tokens.
4. Use read-only analysis first, then gated patch proposals; require independent tests and approval before any merge or deployment.
5. Revisit Google model IDs, pricing, usage restrictions and safety documentation when Carbon becomes accessible.

For buyers, the decision today is whether to monitor a restricted launch or run a controlled pilot on available models. For security teams, it is whether a vetted early-access application is appropriate. For everyone else, Carbon's announced benchmark gains are a reason to prepare a test plan, not to promise a production migration date. Keep your current model and approval controls until the provider confirms the exact access path and your own evaluation supports a change.