$ cat wiki/models/gemini-4-carbon.md
Gemini 4 Carbon
Spec
| Attribute | Value |
|---|---|
| Developer | Google DeepMind (source) |
| Released | not yet — general availability; limited defender rollout planned for announcement week (source) |
| Announced | 2026-10-28 (source) |
| Context window | unknown |
| Pricing | Planned introductory $2/M input · $10/M output; cached input 95% off input price. After introductory period: $4/M input · $20/M output; expiry date unspecified (source) |
| License | unknown |
| Availability | Initial cyber-defender cohort, then additional trusted Fairwind Program partners; broader developer, business and consumer access forthcoming (source) |
Release Date
Google announced Carbon on 2026-10-28, saying the initial rollout would start that week. The announcement does not establish a public endpoint or a general-release date. Context capacity and license terms are not stated in the captured article. (source)
Benchmarks
Google reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench, explicitly evaluated without guardrails. Its claim of 59 on Artificial Analysis Intelligence Index v5.0 is vendor-reported; the live leaderboard was not independently fetched. These scores do not establish performance with deployment safeguards enabled. (source)
Google also reports 0.0% successful cheating and 0.4% cheating attempts on Terminal-Bench 4.0, plus 0.0% attack success on Gray Swan's Indirect Prompt Injection benchmark. These are different measurements, not general guarantees of safe behavior. Chart images and the linked safety report were not independently reviewed. (source)
Use Cases
The announced focus is sustained software engineering and cybersecurity defense: finding vulnerabilities, validating exploitability and patching root causes. Google says internal teams have seen runs lasting days; the article does not quantify ordinary task completion rates or costs. (source)
Compared To
Carbon builds on Gemini 4 Argon. Google reports more than 2x uplift in secure-code generation success on internal benchmarks and approximately 2x fewer unsupported claims. The announcement does not supply enough evaluation detail to infer the same gains on another team's workload. (source)
Sources
- Google announcement, 2026-10-28, captured 2026-10-29. (source)