Google Gemini 4 Carbon is Google’s newest frontier AI model, announced on October 28, 2026. Google calls it its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities. You almost certainly cannot use it yet. Google is giving it first to a subset of cybersecurity defenders, then to vetted partners, and it has set no date for general release.
That gap between announcement and availability is the whole story of Carbon so far. It is a real model with published benchmark claims, not a rumor. But it is also a restricted model, and the benchmark numbers come from Google itself.
What Gemini 4 Carbon is, in plain terms

Gemini 4 Carbon is the second model in Google’s Gemini 4 series, following Gemini 4 Argon. The name follows the same pattern: Argon and Carbon are both chemical elements. Carbon is not a smaller variant, a fine-tune, or a security-only tool. Google presents it as a general frontier model with two advertised strengths: cybersecurity work and long-horizon tasks.
The announcement came from Koray Kavukcuoglu, senior vice president at Google DeepMind and Google’s chief AI architect. His framing sentence called Carbon “our strongest model with state-of-the-art cybersecurity and long horizon capabilities.” That is a claim about capability, and it is worth reading as a claim, not as an independent verdict.
For most readers, the practical definition is simpler. Carbon is Google’s latest top-tier AI model, built to handle complex software and security work over long stretches, and it is currently in a limited rollout while Google refines its safety guardrails. If you use the Gemini app today, you are not using Carbon.
Who can use it right now, and who cannot?

Access starts with a subset of cyber defenders, and it widens from there in stages. Google says Carbon goes first to that defender group, then to additional trusted partners through its Fairwind Program, Google’s vetted early-access scheme. Developers, businesses, and consumers get the model “as soon as possible,” in Google’s words, after feedback and further guardrail work.
As of early November, that means the following. Carbon is not in the Gemini app on any plan, free or paid. There is no public API model ID for it, and Google’s pricing page lists Argon, not Carbon. It is not in coding tools such as Cursor or Copilot. The Fairwind Program page still described Argon and did not mention Carbon as of October 29.
Google is also taking part in the US government’s voluntary process for pre-release model testing. That process is one reason for the staged rollout: Google says it wants feedback and stronger guardrails before a wider release. The company has not published a general release date, and any date you see quoted for subscribers is an inference from Argon’s rollout, not a Google announcement.
That inference keeps appearing, so it is worth explaining. Argon went to Fairwind security teams and then reached paid Gemini app subscribers about two weeks later, on October 13. If Carbon followed the same clock, paying subscribers might see it in the first half of November. But that is one data point, and Google has explicitly said it is still working on guardrails. Treat any November date as a guess, not a promise.
What Google claims it can do

Google’s announcement leads with security benchmarks, and every number below is Google’s own claim, not an independently verified result. Google says the cybersecurity benchmarks were evaluated without guardrails in place, which is a detail worth noting: the scores describe raw capability, not the behavior of the guarded product most users would eventually meet.
On SEC-bench Pro, a test of finding real vulnerabilities in software, Google reports 88.2 percent. On ExploitGym, which measures turning a flaw into a working exploit, Google reports 54.8 percent. On Collinear’s CWE-bench, a test of producing code fixes for vulnerabilities, Google reports 71.0 percent. Collinear’s public leaderboard, dated October 28, does list Carbon first on CWE-bench at 71 percent, ahead of a three-way tie at 68 percent.
Outside security, Google reports that Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index, a composite of independent tests, with a score of 59, one point ahead of Claude Opus 5.5 on that index. On WebDev Arena, a public ranking of models that build websites, Carbon gained 89 Elo points to reach 1768, placing it fourth.
Two cautions apply to all of this. First, benchmarks do not tell the whole story, and frontier scores in these tests sit in the same neighborhood as each other. A first-place ranking on one composite index is not proof of superiority in real work. Second, Google has published a Carbon performance table, but its model-card index still lists Argon, and the safety report linked from the Carbon announcement covers Argon, not Carbon. Independent evaluation has not yet caught up with the announcement.
How Carbon compares with Gemini 4 Argon

Argon is the closest comparison, because it is the model Carbon directly follows. Here is how they line up on what is publicly known:
- Announcement: Argon was announced in late September 2026 and reached paid Gemini app subscribers on October 13. Carbon was announced on October 28.
- Position: Argon was Google’s frontier model for that moment. Carbon is now presented as Google’s strongest model.
- First access: Both models went first to cyber defenders through the Fairwind Program. Carbon is still in that stage.
- General availability: Argon reached subscribers, the paid API, and coding tools within weeks. Carbon has no general availability date.
- Focus: Argon was a general frontier model with strong security capabilities. Carbon pushes further on cybersecurity and on long-horizon work.
- Benchmarks: Carbon’s headline security scores are Google-reported and higher than the figures Google emphasized for Argon, but they come from different benchmark versions and should not be read as a clean head-to-head.
The pattern to watch is Argon’s rollout. Argon moved from defender-only access to subscribers, the API, and coding tools in stages over a few weeks. Carbon may follow the same path, or it may take longer, because Google has tied its wider release to guardrail work and voluntary pre-release testing. The Argon timeline is a guide, not a schedule.
What does long-horizon actually mean?

Long-horizon is a term that sounds technical, but the idea is simple. Most AI models handle short tasks well: answer a question, write a snippet, summarize a document. A long-horizon task stretches over many steps and a long time, sometimes days. Think of a software project that needs planning, writing code, testing it, finding bugs, fixing them, and checking the result, all without losing track of the original goal.
Google says Carbon is built for that kind of work, especially in software engineering and cybersecurity. A security team might ask it to find vulnerabilities across a large codebase, validate which ones are real, and write fixes, as one connected job rather than a series of separate prompts. That is the cyber-defense lifecycle: find, validate, fix.
Whether Carbon actually sustains that kind of work in practice is unproven outside Google. Long tasks are where models tend to drift, lose context, or make confident mistakes. That is also why Google is starting with vetted defenders rather than the general public: the users most able to catch mistakes get the model first, and their feedback shapes the guardrails.
What the announced API pricing means
Google has announced API pricing for Carbon, even though the model is not publicly available through the API yet. The introductory rate is 2 dollars per million input tokens and 10 dollars per million output tokens, matching Argon’s introductory pricing. Cached input tokens are priced 95 percent below the standard input rate. After the introductory period ends, prices rise to 4 dollars per million input tokens and 20 dollars per million output tokens.
Two things to note. First, Google has not said when the introductory period ends, so the higher rate has no start date. Second, because the public API documentation still lists Argon rather than Carbon, the announced Carbon price is a stated future price, not evidence that you can buy API access today.
Google’s push into new AI infrastructure is part of this picture too. The company’s work on specialized hardware, including its Google TPU in Orbit Suncatcher project, shows how much of the AI race now runs on computing capacity, not just model design.
What to watch next
Three things will tell you whether Carbon is becoming a real product or staying a restricted announcement.
First, the Fairwind Program page. When it starts mentioning Carbon instead of only Argon, the wider partner rollout has begun. That page is the only door for security teams that want access now.
Second, the Gemini app and the API documentation. When Carbon appears as an available model, or when Google announces subscriber access, the general rollout has started. Until then, any claim that Carbon is publicly available is wrong.
Third, independent evaluation. When outside researchers reproduce or challenge Google’s benchmark claims, especially the 88.2 percent SEC-bench Pro figure, the picture will sharpen. Vendor benchmarks are a starting point, not a verdict.
What is Google Gemini 4 Carbon: quick answers
What is Google Gemini 4 Carbon?
Google’s newest frontier AI model, announced October 28, 2026, with a focus on cybersecurity and long-horizon tasks.
Can I use Gemini 4 Carbon now?
Almost certainly not. It is available only to a subset of cyber defenders, with wider partner access to follow. Google has set no general release date.
How is it different from Gemini 4 Argon?
Carbon follows Argon as the second Gemini 4 model, with stronger claimed cybersecurity and long-horizon performance. Argon reached paid subscribers on October 13; Carbon has no equivalent date yet.
What are its benchmark scores?
Google claims 59 on the Artificial Analysis Intelligence Index, 88.2 percent on SEC-bench Pro, 54.8 percent on ExploitGym, and 71.0 percent on CWE-bench. These are vendor claims, not independently verified results.
How much will it cost?
Announced API pricing is 2 dollars per million input tokens and 10 dollars per million output tokens as an introductory rate, rising to 4 and 20 dollars later. Google has not said when the introductory period ends.



