On October 28, 2026, Google used its official blog to announce Gemini 4 Carbon, which it calls its strongest AI model yet. Google is pitching Carbon on two things in particular: cybersecurity work and very long, complicated tasks. The catch is that you almost certainly can’t use it today. Starting this week, Google is giving it to a subset of cyber defenders (security teams), and it hasn’t said when regular people will get it.
Here’s what Carbon is, why Google is leading with security, what it costs, and what’s still unproven (spoiler: quite a lot). Google’s official announcement is the primary source for Google’s claims below; third-party sources are linked where we use them. All sources were checked as of October 28, 2026.
What Happened
Here’s the short version: Gemini 4 Carbon is Google’s newest frontier model, which is the industry’s term for a company’s most advanced, top-of-the-line AI. Google says it builds on Gemini 4 Argon, which launched only about two weeks earlier. That’s a fast turnaround, even by AI-industry standards. If you just got used to the name “Argon,” we feel your pain.
According to Google’s announcement, Carbon’s main selling points are:
- Cybersecurity skills: Google says Carbon is very good at finding security holes in software, checking whether those holes can actually be exploited, and writing fixes.
- Long-horizon persistence: Google says Carbon can keep working on one hard problem for a long time. Internally, Google says it has seen the model work for days on difficult engineering problems without giving up. That claim comes from Google’s own internal testing and hasn’t been independently verified from outside the company.
- Safer code than Argon: Google says that on its internal benchmarks, Carbon achieves more than 2x Argon’s success rate at generating secure code. Google hasn’t named that benchmark or published its methodology.
- A #1 ranking on an independent index: Google says Carbon ranks #1 on the newly updated Artificial Analysis Intelligence Index v5.0 with a score of 59. Artificial Analysis itself logged a new evaluation of “Gemini 4 Carbon (High)” on October 28 and published an article headlined “Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5.”
- A deliberate, phased rollout: Google is starting with a subset of cyber defenders this week and will gradually phase in more trusted partners through its Fairwind Program. It says developers, businesses, and consumers will follow “as soon as possible.”
9to5Google’s coverage of the same announcement is useful if you’d like a second read.
What “state-of-the-art cybersecurity” actually means
“State-of-the-art” is a phrase that gets thrown around a lot, so here’s what Google is actually claiming. It points to three security tests, and each one measures a different part of a security team’s job. Think of it like a home inspector who has to find the crack in the foundation, decide whether it’s actually dangerous, and then patch it.
| Test (Google’s own figures) | What it roughly measures | Carbon’s claimed score |
|---|---|---|
| SEC-bench Pro | Finding security holes in software | 88.2% |
| ExploitGym | Working out whether a hole can really be exploited | 54.8% |
| CWE-bench | Writing code that fixes common types of security flaws | 71.0% |
Source: Google’s announcement. Independent third-party reproductions of these exact numbers weren’t found in our research as of October 28, 2026.
The ExploitGym score is the one worth noticing. Just over half sounds modest, and it is. Proving that a vulnerability can really be used in an attack is much harder than spotting one. (It’s the difference between noticing that a window is unlocked and actually climbing through it.)
What “long-horizon” means
Most AI chatbots work in short bursts. You ask, they answer, and if the problem is hard, they often stall or give you something half-finished. When Google calls Carbon “long-horizon,” it means the model is built to keep going through dead ends on a multi-step problem instead of quitting early. Picture a stubborn bug that takes a human engineer several days to fix. Google’s claim is that Carbon can keep working on that kind of bug for days at a time.
That’s an impressive claim. Keep in mind that it comes from Google’s own internal experience, not from an outside test. No independent lab has confirmed this ability outside of Google’s own reporting.
Why Google is leading with security
Google’s rollout plan explains a lot here. A model that’s good at finding and testing software weaknesses is useful to defenders, the people patching their own systems. That same skill could also be useful to attackers. So Google is putting it in the hands of “cyber defenders” first, meaning security teams that use AI to find and patch vulnerabilities. Google describes this as responsible deployment. In plain terms, it wants the good guys to get a head start.
Who Gets It, and When
This is the part most readers care about, and the answer is a little disappointing. Google describes a phased approach rather than a fixed schedule:
- A subset of cyber defenders gets access starting this week.
- Additional trusted partners are phased in gradually through Google’s Fairwind Program.
- Developers, businesses, and consumers get it through Google’s products after Google gathers more feedback and iterates on its guardrails.
For that last step, Google’s only wording is “as soon as possible.” That’s not a date. There’s no confirmed timeline for when Carbon will appear in the regular Gemini app or any other consumer product. Google’s announcement doesn’t describe any general sign-up for consumers.
Some outside sites, like projedefteri.com, track possible release dates. Treat those as guesses until Google says otherwise.
What It Costs
Google has published introductory API pricing. An API is the way software developers plug an AI model into their own apps, so this is a price for businesses and developers, not a Gemini app subscription.
| Item | Introductory price (per Google) |
|---|---|
| Input tokens | $2 per million |
| Output tokens | $10 per million |
| Cached input tokens | 95% off the input rate |
| After the introductory period | $4 per million input, $20 per million output |
Source: Google’s announcement.
A token is a small chunk of text, roughly a piece of a word. “Input” is what you send the model, and “output” is what it writes back. “Cached” input is text the model has already seen and can reuse, like the same long instructions sent over and over. Applied to the $2 introductory input rate, a 95% discount brings cached input down to about 10 cents per million tokens.
Google says this is the same introductory pricing as Gemini 4 Argon. So if you’re a small business already budgeting for Argon through the API, Google isn’t charging more for the new model, at least for now. Those are introductory rates, though.
Google says that once the introductory period expires, Carbon will cost $4 per million input tokens and $20 per million output tokens, double the launch price. It hasn’t said when the introductory period ends. Google also hasn’t announced any consumer pricing or a new subscription tier for Carbon, so don’t assume one exists.
How It Stacks Up
Gemini 4 Carbon vs. Gemini 4 Argon
Google says Carbon builds on Gemini 4 Argon, which has only been around for about two weeks. According to Google, Carbon:
- Achieves more than 2x Argon’s success rate at generating secure code on Google’s internal benchmarks (Google hasn’t named the benchmark or published its methodology).
- Holds on to long, difficult tasks better than Argon.
One thing to watch: benchmark numbers circulating online for tests like DeepSWE v1.1 and LVBench aren’t always clearly labeled by model version. We couldn’t independently confirm which figures belong to Argon versus Carbon in every secondary roundup. When comparing scores from unofficial sources, check carefully which model each number actually describes.
Google also shared a WebDev Arena result. Carbon gained 89 Elo points (a chess-style rating) to reach #4 at 1,768. Google highlighted that as a big improvement, but it didn’t claim the top WebDev Arena spot.
Gemini 4 Carbon vs. Claude and ChatGPT
This is the question everyone asks: is it better than ChatGPT or Claude? The short answer is that Carbon now tops one independent composite index, and there’s no head-to-head data yet for everyday tasks.
The one cross-vendor comparison from an independent tester is the Artificial Analysis Intelligence Index v5.0, which rolls many evaluations into a single score. Google’s announcement says Carbon scores 59 there, and Artificial Analysis published an Oct. 28 article titled “Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5”:
| Model | Artificial Analysis Intelligence Index v5.0 | Where the number comes from |
|---|---|---|
| Gemini 4 Carbon (Google) | 59 (#1) | Google’s announcement; Artificial Analysis’s Oct. 28 “new #1” article |
We haven’t independently confirmed the current scores for Claude or ChatGPT models on that index, so we’ve left them out of the table rather than guess at the margin. Google’s headline security tests (SEC-bench Pro, ExploitGym, CWE-bench) come from Google’s own announcement, so those scores are still Google’s own claims until outside testers reproduce them. Some non-official compilations, such as kie.ai, also circulate coding-test numbers for rival models, but none come from the model makers, so we’ve left them out.
The Reaction
Reaction to the official announcement is still forming, since it went out on Oct. 28.
Most of the online chatter so far predates the announcement. Before launch, Nokia Power User reported what it described as a Carbon checkpoint appearing on an AI arena; Google hasn’t confirmed that. Early tester write-ups, such as TestingCatalog’s, shared general impressions of unreleased builds without describing the builds, workloads, or conditions.
Those are anecdotes from unreleased test versions, not measured results, so don’t treat them as facts about the shipping model.
Our Take
Who should care right now:
- Security teams: Google designed Carbon for you, tuned its benchmarks for your work, and put you first in line. If your organization does vulnerability research or patching, it’s worth finding out whether you qualify for early access or the Fairwind Program.
- Engineering teams with long, stubborn problems: If past AI tools have given up on your multi-day bugs, Carbon’s long-horizon pitch speaks directly to you. Wait for broader developer access and independent testing before you reorganize around it.
- Small businesses budgeting for AI: The introductory API pricing matches Argon’s, but budget for the $4/$20 standard rate once the introductory period ends.
Who should wait: Pretty much everyone else. If you use Gemini, ChatGPT, or Claude to draft emails, plan trips, or settle trivia arguments, nothing changes for you today. There’s no Carbon in your app and no release date. Also, Carbon’s best claimed skills, like finding exploits, aren’t things most of us do on a Tuesday afternoon.
Our view: the cautious rollout is the most interesting part of this announcement. Google is saying a model this good at security work shouldn’t be handed to everyone on day one. Artificial Analysis’s #1 listing is a good early sign, but Google’s security benchmark scores still need independent reproduction.
What to Do Next
If you’re a regular user, your one practical step is to bookmark Google’s official announcement and check it for rollout updates instead of trusting release-date rumors.
Also, AI tools are getting better at finding software weaknesses, on both the defending and attacking sides. That makes basic security habits a little more important. Keep your devices updated, use a password manager, and consider a hardware security key for important accounts like email and banking. None of that needs Carbon, and all of it helps right now.
Wrapping Up
Gemini 4 Carbon is Google’s big claim that it has built a much stronger AI, especially for finding and fixing security holes and for sticking with hard problems for days. For now, it’s mostly available to security teams. Artificial Analysis now ranks it #1 on its Intelligence Index, but its headline security scores are still Google’s own figures. And there’s no date yet for when regular people can try it.
Honestly? That’s fine. A careful rollout for a security-focused model makes sense, and it gives outside testers time to check Google’s numbers before the hype takes over.
| What to know | Details | Applies To |
|---|---|---|
| Announced | October 28, 2026, on Google’s official blog | Everyone |
| Who gets it first | Cyber defenders, then Fairwind Program partners | Security teams, partners |
| API pricing | $2/M input, $10/M output introductory ($4/$20 afterward), 95% off cached input | Developers, businesses |
| Consumer release | No date (“as soon as possible”) | Everyday users |
| vs. Claude/ChatGPT | #1 on Artificial Analysis Index v5.0 (score 59); security scores are Google’s own | Everyone |