Google announced Gemini 4 Carbon on Wednesday, October 28. It is the company’s new frontier model, and it arrives just two weeks after Argon. On top of that, Google says it is its strongest model yet for cybersecurity and long, multi-step coding work.
So should you care? In short, yes, but you can’t use it yet. In this post, I will walk you through the benchmarks, the safety claims, the price and the limited rollout. Then I will share how I would plan for it.
What Is Gemini 4 Carbon?
First, some context. In its launch post, Google calls Carbon its next frontier model. It builds on Gemini 4 Argon, which I covered in my Gemini 4 Argon launch post.
In other words, Carbon is not a small update. Google aimed Gemini 4 Carbon at hard, long-horizon engineering problems. For example, the company says Carbon has run for days on a single long task.
General skills improved as well. For instance, Carbon gained 89 Elo points on WebDev Arena and now sits at #4 there.
Gemini 4 Carbon Benchmarks: Number One, With Caveats
Now let’s look at the numbers. Keep in mind that most of them come from Google itself.
The headline result comes from a third party, though. Carbon scores 59 on version 5 of the Artificial Analysis Intelligence Index. That puts it at #1, ahead of Claude Opus 5.5 and GPT-6 Astra, as The Verge also reports.
However, the full picture is more mixed. According to Artificial Analysis, Carbon leads 4 of the 9 index tests. Still, it trails Claude Opus 5.5 on coding agents, scoring 59 against 62. Similarly, it ranks second on Terminal-Bench 4.0 at 64%.
Knowledge work tells a similar story. Carbon gained 126 Elo points over Argon on AA-Briefcase. Even so, models like Grok 4.7, Muse Wonder and Kimi K3.1 Pro still beat it there.
There is one more catch. Artificial Analysis says Carbon’s hallucination rate rose from 4% with Argon to 11%. That is still the second lowest of any model it tracks. By contrast, it measured 50% for both Claude Opus 5.5 and GPT-6 Astra.
Why Cybersecurity Is the Real Story
Here is where Gemini 4 Carbon stands out. Google reports strong scores on three key security benchmarks. Notably, Google ran these tests without guardrails.
- SEC-bench Pro, for finding real vulnerabilities: 88.2%
- ExploitGym, for checking if a bug is exploitable: 54.8%
- CWE-bench, for patching the root cause: 71.0%
Think of these as three steps. First, you find the bug. Next, you check if attackers can use it. Finally, you write a working fix. Carbon covers the whole chain.
Google adds some internal examples too. For instance, it says Carbon more than doubled Argon’s success rate at writing secure code. Moreover, the Chrome team found Carbon over 10 times better at high-quality patches than another lab’s large model.
Of course, a model that finds and exploits bugs can also help attackers. That is exactly why Google is so careful with access.
Safety Claims: Fewer Shortcuts and Less Bluffing
Google spent a lot of the launch post on alignment. It says Carbon uses the same safety stack as Argon. That stack combines training, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring.
Then Google lists the gains. For example, successful cheating on Terminal-Bench 4.0 dropped to 0.0%. Cheating attempts fell to 0.4%. In addition, high-severity misalignment alerts dropped by almost 80%.
Honesty improved too, at least by Google’s measure. Carbon makes about half as many unsupported claims as Argon. Also, overreach fell 26% and under-service fell 29%.
Prompt injection is the last big claim. Argon had a 0.7% attack success rate on Gray Swan’s Indirect Prompt Injection benchmark. Meanwhile, Carbon reached 0.0%.
Still, remember that these are self-reported results. So I would treat them as promising, not proven, until outside testers confirm them.
Gemini 4 Carbon Pricing and Who Can Use It
Access is the biggest limit right now. Starting this week, Google is giving Carbon to a small group of cyber defenders. After that, it will slowly add trusted partners through its Fairwind Program. Wider access comes later, with no public date yet.
Google also says it is part of the U.S. government’s voluntary process for testing models before release. In short, this is a slow, careful rollout.
Pricing is already set, though. Here is how it works:
- Introductory price: $2 per million input tokens and $10 per million output tokens
- Standard price after the intro period: $4 input and $20 output
- Cached input: 95% off the input price
Furthermore, Artificial Analysis puts Carbon’s cost at $7.76 per index task at standard prices. That is about 23% less than Claude Opus 5.5, but about 1.2 times the cost of GPT-6 Astra.
What I Would Do Next
So what does this mean for you? Here is how I would plan, step by step.
- First, if you work in security, check whether your company can join the Fairwind Program.
- Second, keep building on Argon for now, so you are ready to compare the two later.
- Next, list your long, painful engineering tasks, such as bug hunts or big refactors. These are the best tests once access opens.
- Finally, run your own evals before you switch. Benchmarks are a starting point, not a verdict.
Overall, Gemini 4 Carbon shows how fast the frontier now moves. Two weeks after Argon, Google is back at the top of a major index. Even so, the real test comes when developers can use it. Until then, I would watch, prepare, and wait for independent results.
No comment