Google has launched Gemini 4 Carbon — not for developers, not for consumers, and not even for most enterprises. It’s live only for a narrow cohort: cyber defenders, starting this week. No public API. No Android or web app integration.
No pricing for end users. Just a $2-per-million-input-token API tier, locked behind Google’s Fairwind Program and reserved for security teams vetted by the company. The model is real. The access is not open.
And the timing is pointed: two weeks after Gemini 4 Argon, and just as US federal agencies begin pre-release AI model testing under new voluntary guidelines. <br><br>Carbon is billed as Google’s most capable model for long-horizon tasks, meaning it’s built to run for hours or days on complex engineering problems, not just answer questions. In Google’s internal tests, it diagnosed, root-caused, and patched subtle bugs across massive codebases without human interruption.
Its benchmark scores, 59 on the Artificial Analysis Intelligence Index v5.0, 1768 on WebDev Arena, are self-reported and unverified by third parties. So are its security claims: zero percent attack success rate on Gray Swan indirect prompt injection testing (down from 0.7% for Argon), and an 80% drop in high-severity misalignment alerts during live coding. <br><br>The Chrome security team says Carbon produced high-quality patches at more than ten times the rate of another lab’s large frontier model, again, internally measured, no external validation.
Google also says Carbon outperformed Argon by more than 2× in generating secure code in internal benchmarks. All of that matters only if you’re among the first wave of defenders granted access. Everyone else waits. <br><br>Pricing is set, but not yet usable. Introductory rates, $2 per million input tokens, $10 per million output tokens, mirror Argon’s launch pricing.
Cached inputs cost 95% less. After the introductory period ends, those jump to $4 and $20. Google hasn’t said when that happens. Nor has it given a date for broader developer or consumer access.
It only says it’s “gathering feedback and iterating on guardrails” and will open up “as soon as possible.” <br><br>This isn’t a general release. It’s a controlled rollout, one that treats AI model access like a security clearance. For now, Gemini 4 Carbon exists only where it’s needed most: inside threat-hunting labs, incident response war rooms, and government-adjacent red teams.
Not in your phone. Not in your browser. Not in your next-gen Android app. <br><br>Next step? Fairwind Program participants begin onboarding this week.
Everyone else watches the benchmarks, reads the blog post, and waits for Google to decide what ‘as soon as possible’ actually means.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.