Google unveiled Gemini 4 Carbon on Wednesday, October 28, 2026 — its latest frontier AI model, and the second new Gemini in a fortnight. It arrives just two weeks after Gemini 4 Argon, and it has a much narrower job description than a general-purpose flagship: coding and cybersecurity, done properly, over long and complicated jobs.
That focus is what makes Carbon interesting. It is also why most developers reading this cannot actually try it yet. Let's separate the two halves of the story: what Google says the model can do, and who can actually touch it today.
What launched
Carbon is a frontier model in the Gemini family, aimed at developers and enterprise customers. Google is positioning it directly against OpenAI and Anthropic in the race for more capable coding and cybersecurity systems — the part of the AI market where a model that writes plausible-looking code is no longer enough, and a model that writes secure code is the whole point.
Google says Carbon delivers state-of-the-art performance across cybersecurity tasks, including vulnerability discovery, exploitation, and software patching. It is also designed for complex, long-running engineering and coding tasks — the kind of work where a model has to hold a large codebase in mind, make a series of dependent changes, and still be coherent at the end.
The one hard number Google has given so far is striking, and it deserves a careful reading: Carbon generates secure code at more than twice the success rate of Gemini 4 Argon on Google's internal benchmarks.
"More than twice" sounds like a marketing line until you unpack what it actually claims. It is a comparison against Argon, on Google's own internal tests, for secure code specifically — not raw code output, and not a public benchmark you can reproduce yourself. The direction of travel is clear, and the size of the jump is large. But it is Google grading its own homework, on tests it designed, and the details of those tests have not been published. Treat it as a strong signal, not a verified fact about how Carbon will behave on your codebase.
Carbon vs Argon: what actually changed
Argon, introduced two weeks ago, expanded the Gemini family for developers and enterprise customers. Carbon builds on that same family, but with a specialist brief.
Think of it this way:
- Argon was the broader developer-and-enterprise release — the new generation arriving.
- Carbon is the security-and-engineering specialist within that generation — tuned for the work where a bug is not just an inconvenience but a vulnerability.
The pricing tells the same story. Carbon is priced at $2 per million input tokens and $10 per million output tokens, matching Gemini 4 Argon's introductory pricing. In plain terms: if you feed it a million tokens of code and context (roughly a very large codebase review, depending on how you count it), the input side costs $2; every million tokens it writes back costs $10. Output is where the bill lives, which is worth remembering for long-running tasks — a model designed to work for hours will generate a lot of output, and that is priced five times higher than input.
That pricing parity matters for a second reason. Google is not charging a specialist premium over Argon, at least at launch pricing. If and when Carbon reaches wider availability, teams already budgeting around Argon pricing will not need a new pricing model to evaluate it. Whether that introductory pricing holds is, of course, a different question — and one Google has not answered yet.
Who can actually use it today
Here is the part most headlines will bury, and it is the most important detail in the whole announcement: Carbon is being made available first to a subset of cybersecurity defenders through Google's Fairwind Program.
Not to all developers. Not to all enterprise customers. Not through a general API signup you can complete this afternoon. A subset of cybersecurity defenders, through a specific program, while safeguards are still being tested.
That phrasing — safeguards still being tested — is doing a lot of work, and it is worth taking seriously rather than treating it as boilerplate. A model that is genuinely good at vulnerability discovery and exploitation is, by definition, dual-use. The same capability that finds a flaw in your code before an attacker does can find a flaw in someone else's code for an attacker. Google is saying, in effect, that it knows this, and that it is limiting initial access to defenders while it tests the guardrails.
That is a reasonable approach, and it is also a reason to be sceptical of any coverage that talks about Carbon as if it were already a tool you can drop into your pipeline. Right now, for most developers, Carbon is something to watch and plan around — not something to build on.
If you are in the defender community, the practical question is whether you are in that Fairwind subset, and what the program's terms look like. If you are not, the practical question is what changes when wider access arrives: pricing is already public, the capability claims are on the record, and the gap between announcement and availability is now the thing to measure.
Why a developer or security engineer should care
Even with limited access, Carbon is worth your attention for three reasons.
First, secure code generation is becoming the benchmark that matters. For two years, coding models have been judged mostly on whether their code works. Carbon's headline claim is about whether its code is secure — a harder, more useful test. If that becomes the standard way frontier coding models are compared, it changes what "good at coding" means, and it puts pressure on every other provider to publish equivalent numbers.
Second, long-running engineering work is the next bottleneck. Short, single-file tasks are largely solved. The painful work in real engineering is multi-step: refactor across a codebase, patch a vulnerability and its knock-ons, migrate a service without breaking its dependents. A model explicitly designed for complex, long-running tasks is aimed at exactly that pain — and if it delivers, it changes how much of that work a small team can take on.
Third, the defender-first rollout is a template to watch. How Google decides who counts as a defender, how it tests safeguards, and when it widens access will set expectations for every similarly capable model that follows. OpenAI and Anthropic are competing in the same space, and their access policies will be judged against this one. The Fairwind Program is not just a distribution channel; it is a policy statement about who gets powerful security capabilities first, and why.
What to watch next
- Wider availability: when, if at all, Carbon moves beyond the Fairwind subset — and whether pricing stays at $2/$10 when it does.
- Independent evaluation: third-party tests of secure-code generation and vulnerability work, on public benchmarks, rather than Google's internal ones. The "more than twice Argon" claim needs an outside check before it should drive a buying decision.
- Safeguard details: what Google's testing actually covers, and what it finds. "Safeguards still being tested" is an honest admission; the follow-up is what the tests show.
- Competitive response: how OpenAI and Anthropic answer a security-specialist frontier model, and whether they match the defender-first framing or go straight to general availability.
What we don't know yet
Honesty requires a list, because the announcement leaves several important questions unanswered:
- General API availability: Google has not said when — or whether — Carbon will be generally available beyond the Fairwind subset. Do not assume you can sign up for it.
- Context window: no context-window figure has been announced. Any article quoting one for Carbon is guessing.
- Public benchmark scores: beyond the internal "more than twice Argon" claim for secure code, no public benchmark numbers have been given. There is nothing yet to compare against OpenAI or Anthropic models on an independent test.
- Fairwind eligibility: the criteria for the subset of defenders getting first access have not been spelled out publicly.
- How "state-of-the-art" was measured: the cybersecurity performance claim rests on Google's own evaluation so far, not on an independent audit.
None of that makes Carbon unimportant. A frontier model that is genuinely better at finding vulnerabilities, patching them, and writing secure code over long tasks would be a serious tool — for defenders first, and eventually for the rest of us. But the honest status today is: announced, priced, narrowly available, and still being safeguard-tested. The model to watch is real. The model to use, for most of us, is not here yet.
Disclosure: This article is based on Google's announcement of Gemini 4 Carbon on October 28, 2026, as reported by Reuters. We have not tested Carbon hands-on — it is initially available only to a subset of cybersecurity defenders through Google's Fairwind Program. No affiliate links appear in this article, and no benchmark, context-window, or availability claim beyond the announcement has been added.
Sources
- Reuters, "Google unveils Gemini 4 Carbon frontier AI model," October 28, 2026
Top comments (0)