What Is Gemini 4 Carbon? Price, Access and Benchmarks
The short answer
Gemini 4 Carbon is Google DeepMind’s new top model, announced October 28, 2026, aimed at cybersecurity and long-running engineering tasks — but almost nobody can use it yet. It launches to a small group of cyber defenders first; general access has no date. When it opens, Google says it will cost the same introductory $2 / $10 per million tokens as Gemini 4 Argon.
| Gemini 4 Carbon | Gemini 4 Argon | |
|---|---|---|
| Announced | October 28, 2026 | September 30, 2026 |
| General availability | Not yet — cyber defenders and Fairwind partners first | October 13, 2026 |
| Intro price (input / output per 1M) | $2 / $10, cached input 95% off | $2 / $10 through Nov 30, 2026 |
| Price after intro | $4 / $20 | $4 / $20 from Dec 1, 2026 |
| Artificial Analysis Index v5.0 | 59 (high), #1 | 50 |
Index scores from Artificial Analysis v5.0 as reported October 28, 2026; v5.0 is not comparable with earlier index versions. See our current API prices.
What Google claims Carbon does better
Cybersecurity across the full lifecycle. Google reports state-of-the-art results on finding vulnerabilities (SEC-bench Pro 88.2%), validating them through exploitation (ExploitGym 54.8%) and patching root causes (Collinear’s CWE-bench 71.0%), all evaluated without guardrails. Internally it reports more than 2× uplift in generating secure code over Argon, and the Chrome team reported Carbon producing high-quality patches more than 10× better than “another lab’s large frontier model”.
Long-horizon engineering. Google says internal teams have seen Carbon work for days on multi-step tasks such as root-causing subtle bugs across large codebases. Web development improved by 89 Elo to 1768 on WebDev Arena, fourth place.
Fewer shortcuts and fabrications. Google reports successful cheating on Terminal-Bench 4.0 cut to 0.0% (attempts 0.4%), high-severity misalignment alerts in live internal coding down almost 80%, and roughly half as many unsupported claims as Argon. Google now reports Terminal-Bench 4.0 scores run by Artificial Analysis rather than its own runs.
Where it does not lead
In Google’s 17-row comparison, GPT-6 Astra still wins FrontierSWE 2 (software engineering) and Claude Opus 5.5 wins OSWorld-2.1 (operating a computer). If your workload is computer use, Opus 5.5 at $4 / $20 remains the stronger documented choice today; for most other work, Carbon is the benchmark leader you cannot yet buy.
Why access is staged
Carbon’s cyber capabilities are the reason. Google says it is taking a “deliberate approach to rollout”, participating in the U.S. government’s voluntary pre-release testing, and gathering feedback on guardrails before wider release. The Fairwind Program is the same trusted-access route Argon used before its October 13 general availability.
What to do now
- Building today: use Gemini 4 Argon, which is generally available at the same introductory price; Carbon should be a model-ID change when it opens.
- Security team: apply through Fairwind if you qualify; Carbon’s vulnerability-finding scores are its strongest claim.
- Choosing a frontier model in November 2026: compare the shipped options in Gemini 4 Argon vs Claude Opus 5.5 vs GPT-6 Astra.
- Budgeting: plan on the $4 / $20 post-introductory price, not $2 / $10.
Last verified: October 29, 2026.