Gemini 4 Carbon is Google’s new flagship AI model, announced on October 28, 2026, but you can’t use it yet. For now, only a small group of cyber defenders has access, and Google hasn’t given a date for everyone else.
Here’s what Google has confirmed, what it costs, and what you can do while you wait.
Quick facts first:
- Announced October 28, 2026, in a Google blog post.
- Access is limited to a subset of cyber defenders through Google’s Fairwind Program.
- Not yet available to developers, businesses, or regular users, and there’s no public timeline.
- Introductory API pricing: $2 per million input tokens and $10 per million output tokens.
- Google says it ranks #1 on the Artificial Analysis Intelligence Index v5.0 with a score of 59.
What Gemini 4 Carbon actually is
Carbon is the newest model in Google’s Gemini 4 family, sitting above the earlier Gemini 4 Argon. Google’s pitch is simple: it’s built to work on hard problems for a long time without losing the thread.
Google says internal teams have watched it run for days on multi-step engineering jobs, like tracking down a bug and patching it. It reportedly keeps going through dead ends instead of giving up.
Google also says the model is better behaved. It claims fewer shortcuts, less made-up detail, and better handling of vague instructions. Treat that as a company claim until independent testers confirm it.
One number backs this up a little. Google says Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0%, and that it blocked every attack on the Gray Swan prompt injection test. Those are Google’s own results, so wait for outside testing before you rely on them.
Why Gemini 4 Carbon is going to cyber defenders first
Carbon is unusually good at finding and exploiting software flaws, and Google clearly knows that cuts both ways. So it’s starting with security teams who can use it to find and fix holes before attackers do.
Google says it will add more trusted partners over time and is still refining its guardrails. It’s also taking part in the U.S. government’s voluntary pre-release model testing process.
This follows a wider trend. Anthropic recently said an unreleased model built working exploit chains across five platforms, which we covered in our exploit chains and patching guide. Labs are now holding their strongest models back until safety work catches up.
Gemini 4 Carbon pricing: cheap at first, then double
Google lists API prices even though most people can’t call the model yet. The introductory rate matches Argon, then it doubles.
Here’s how the numbers compare:
| Period | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Introductory | $2 | $10 |
| After the introductory period | $4 | $20 |
| Cached input tokens | 95% off the input price | Not applicable |
The doubling matters if you plan to build on Carbon. Budget for the later price, not the launch price. For a comparison point, OpenAI recently cut its own rates, as we explained in our GPT-6.1 Terra pricing breakdown.
Gemini 4 Carbon benchmarks in plain English
Google shared a long list of scores. These are the ones that are easiest to understand, and all of them come from Google itself.
| Test | Carbon result | What it measures |
|---|---|---|
| Artificial Analysis Intelligence Index v5.0 | #1, score of 59 | Overall ranking across many tasks |
| SEC-bench Pro | 88.2% | Finding security vulnerabilities |
| ExploitGym | 54.8% | Proving a flaw can be exploited |
| CWE-bench | 71.0% | Writing patches for flaws |
| WebDev Arena | #4, 1768 Elo | Building websites and web apps |
One caveat: Google says the cybersecurity scores were measured without guardrails. Real-world results from the locked-down version may look different.
It’s worth noting that Carbon is only fourth on WebDev Arena. It leads on overall intelligence, not on everything.
What to do while you wait for Gemini 4 Carbon
If you’re a regular user, nothing changes today. Google’s post doesn’t mention the Gemini app, the API, or Vertex AI, so keep using what you have. The Gemini app you use today stays the same.
The model is also a reminder that agentic coding tools are moving fast. A model that can work for days on one task changes how teams plan code reviews, testing, and security checks, so it’s smart to start thinking about those habits now.
If you’re a developer, don’t rewrite anything yet. Keep your code flexible enough to swap models, and watch Google’s blog for a wider release date.
If you work in security, you can look into the Fairwind Program, since that’s the only route in right now.
Frequently Asked Questions
When is the Gemini 4 Carbon release date?
Google announced it on October 28, 2026, but only for select cyber defenders. There’s no public release date for everyone else, just “as soon as possible.”
Is Gemini 4 Carbon free?
Google’s post doesn’t mention a free tier. The only pricing it gives is for the API.
Can I use Gemini 4 Carbon in the Gemini app?
Not yet. Google hasn’t said when, or if, it will come to the Gemini app.
How much does Gemini 4 Carbon cost?
The introductory API price is $2 per million input tokens and $10 per million output tokens. After that period, it rises to $4 and $20.
My take: Carbon looks like a strong model, but the story right now is the cautious rollout, not the benchmarks. Don’t plan around it until you can test it yourself. If the later price holds, it will be a premium choice, so the cheaper Argon is the smarter default for most projects in the meantime.
Leave a Reply