Google announced Gemini 4 Carbon on October 28, 2026, with selected cybersecurity defenders due to receive access that week. The model targets demanding software and security work, but Google has not set a date for general availability.
For teams protecting hospitals, public services or widely used software, the immediate question is whether they can join the early rollout. For everyday Gemini users and small businesses, this is not yet a Carbon upgrade they can switch on.
Who can get access
Google says access will expand gradually through its Fairwind Program. The programme prioritizes governments, critical infrastructure operators and core technology platforms; academic labs focused on defensive benchmarking can also apply.
Applying does not guarantee Carbon access. Fairwind’s published model-specific terms still name Gemini 4 Argon, and Google’s public API catalogue does not yet list a Carbon identifier. Prospective users should confirm Carbon access, permitted uses, data-handling terms and delivery timing before scheduling a project around it.
What the performance evidence shows
Google reports improvements in finding and fixing software vulnerabilities. Separately, benchmark maintainer Collinear’s CWE-bench lists Carbon running with Google’s Antigravity coding tools at a 71% average single-attempt pass rate, against 68% for Argon with the same tools.
The evaluation covers 120 tasks, with four runs per task. To pass its automated checks, a proposed fix must stop the exploit while keeping existing tests passing. The practical implication is narrower than a promise of autonomous repair: the results support considering the model for evaluation, not assuming it can safely patch a business’s systems unattended.
Google says the external cybersecurity results it cites were measured without guardrails, so they should not be read as deployed-partner performance. For a related sign-off question, RohitAI’s guide to AI-assisted Java migrations explains why preserving software behavior matters beyond a successful build.
What teams should budget for
The announced introductory prices are $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 respectively. Tokens are chunks of data used to measure what a model processes and generates. Carbon’s introductory period has no announced end date; Argon’s November 30 deadline should not be assumed to apply.
Illustrative calculation: 10 million uncached input tokens plus 1 million output tokens would cost $30 at introductory rates and $60 at standard rates. This assumes unchanged billed volumes and excludes tools, retries, taxes and other charges. It is a usage example, not an estimate for a complete security scan.
For budgeting, the standard rates provide a firmer planning basis than a discount with no published duration. Access still comes first: an eligible organization can apply to Fairwind, but should obtain explicit Carbon terms before committing to deployment.