# Gemini 4 Carbon

Google, released 2026-10-28

**Verdict: conditional.** Conditional. Gemini 4 Carbon is Google's strongest model by its own account, announced October 28, 2026 with what Google calls state-of-the-art cyber results, but access is limited to a subset of cyber defenders and trusted partners through the Fairwind Program. Every figure is Google's own and we have not tested it. Right for security teams that can qualify; everyone else has to wait for the broader release, which Google has not dated. Last verified 2026-11-06.

## Pricing

- Input: $2 / 1M tokens (intro)
- Output: $10 / 1M tokens (intro)
- Context: Intro rate, then $4/$20; gated, no API id

## Strengths

- Vulnerability discovery, exploitation and patching research by defenders who qualify for access; Google reports SEC-bench Pro 88.2%, ExploitGym 54.8% and CWE-bench 71.0% (Google's claims, untested by us)
- Secure code generation; Google says Carbon delivers more than 2x uplift in success rate over Argon (Google's claim)
- Security agents exposed to untrusted content; Google reports 0.0% attack success on Gray Swan's indirect prompt injection benchmark, down from Argon's 0.7% (Google's claim)
- Anyone outside the defender and Fairwind partner groups; the post says a subset of cyber defenders first, then additional trusted partners, and developers, businesses and consumers only "soon"
- Budgeting beyond the introductory price; the post gives $4 input and $20 output per 1M tokens after the introductory period and states no end date
- Decisions that need independent benchmark confirmation; every figure comes from Google's announcement, including the Artificial Analysis figure of 59, which we did not check against Artificial Analysis
- Planning around context size or an API id; the post gives neither, and the Gemini API pricing page and changelog we read list no Carbon entry

## Benchmarks

- SEC-bench Pro (Google-reported): 88.2% (https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)
- ExploitGym (Google-reported): 54.8% (https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)
- CWE-bench (Google-reported): 71.0% (https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)
- WebDev Arena (Google-reported): #4 (1768) (https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)
- Artificial Analysis Intelligence Index v5.0 (as stated by Google): 59 (#1) (https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)
- Gray Swan indirect prompt injection attack success rate (Google-reported, lower is better): 0.0% (https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)

## Sources

- [Google blog: Introducing Gemini 4 Carbon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)
- [Gemini API pricing](https://aistudio.google.com/docs/pricing)
- [Gemini API changelog (10 and 12 November 2026 entries)](https://aistudio.google.com/docs/changelog)

Canonical: https://neomanex.com/models/gemini-4-carbon