Google announces Gemini 4 Carbon, a frontier model initially limited to cyber defenders

The short version

Google announced Gemini 4 Carbon, a frontier model for long-horizon engineering and cyber defense. It starts with selected defenders and trusted partners before any broader developer, business, or consumer availability.

What changed

Gemini 4 Carbon is a new Gemini frontier model. Google says it improves over Gemini 4 Argon for coding, agentic work, multimodal understanding, and defensive cybersecurity, with a staged Fairwind Program rollout.

What it means for you

This is a new frontier-model release with a security-focused access policy, rather than a routine feature update. Teams evaluating coding or defensive security agents now have a new option to track, but cannot assume general API access or safe autonomous operation.

Google has announced Gemini 4 Carbon, a new frontier model that it positions for difficult, long-running engineering work and cybersecurity defense. The release is material because it is a new model and a deliberately security-focused deployment, not an incremental update to an existing developer tool.

Google says Carbon improves on Gemini 4 Argon across coding, agentic tasks, full-stack web work, multimodal understanding, and scientific tasks. Its announcement highlights benchmark results for finding, validating, and patching vulnerabilities. These are Google-reported launch evaluations, so they indicate the company's claims and should not substitute for an organization's own reliability, security, or cost evaluation.

Access is limited at launch. Google says a subset of cyber defenders receives Carbon first, with additional trusted partners joining through its Fairwind Program. The company says it will gather feedback and iterate on guardrails before offering the model to developers, businesses, and consumers. Teams should therefore not plan an integration on the assumption that a general Gemini API option is already available.

Google lists introductory API prices of $2 per million input tokens and $10 per million output tokens, with cached input 95% below the input price. It says the post-introductory prices will be $4 input and $20 output per million tokens. Pricing and access can change, so procurement teams should check the live official documentation before setting production budgets.

For an early evaluation, use a non-production repository with deterministic tests, require human approval for every patch, and keep tools and credentials least-privileged. A security-oriented model can still misread code, follow hostile text in a ticket, report false findings, or generate a damaging change. Measure end-to-end task success, false-positive rates, review rejection, retries, and total cost before giving an agent broader authority.

Related directory entry

Fact check

VerifiedGoogle announced Gemini 4 Carbon as a new frontier model for long-horizon engineering work and cybersecurity defense.Evidence

VerifiedGoogle says initial access begins with a subset of cyber defenders and expands gradually to trusted Fairwind Program partners before broader developer, business, and consumer availability.Evidence

VerifiedGoogle lists introductory pricing of $2 per million input tokens and $10 per million output tokens, then says $4 and $20 apply after that period.Evidence