Gemini 4 Carbon: Our next frontier model

What it is

Gemini 4 Carbon is a Google DeepMind model announced about two weeks after Gemini 4 Argon. DeepMind reports it ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. It also reports #4 on WebDev Arena at 1769 (+89 Elo). The cyber results are SEC-bench Pro 88.2%, ExploitGym 54.8% and Collinear's CWE-bench 71.0%, all described as run without guardrails.

What the page adds beyond the headline numbers

On internal benchmarks, Carbon has more than 2x Argon's success rate at generating secure code. The Chrome team reports Carbon produced high-quality patches more than 10X better than another lab's large frontier model. The page does not name that model or give the measurement. Wiz says Carbon leads its web penetration testing benchmarks.

Safety figures, all reported by DeepMind:

• Successful cheating on Terminal-Bench 4.0 fell to 0.0%, with cheating attempts at 0.4%.
• High-severity misalignment alerts dropped by almost 80% in internal live coding.
• Unsupported claims fell about 2x against Argon.
• Overreach fell 26% and under-service fell 29%.
• Gray Swan's indirect prompt injection benchmark: 0.0% attack success, against 0.7% for Argon.

DeepMind says internal teams saw Carbon run for days on multi-step tasks such as root-causing and patching bugs in large codebases and low-level systems.

What to do

Access is limited. Carbon is available to a subset of cyber defenders starting the week of the announcement, with further trusted partners added gradually through the Fairwind Program. The page gives no application route or eligibility criteria. General availability to developers, businesses and consumers is described only as "as soon as possible".

Planned introductory pricing matches Argon: $2 per million input tokens and $10 per million output tokens. Cached input tokens cost 95% less than the input price. Teams budgeting for agentic security scanning can use these rates now.

Caveats

• Every figure comes from DeepMind. The page links no methodology or comparison scores for other models on the three cyber benchmarks, so "state-of-the-art" cannot be checked from it.
• The cyber benchmarks were run without guardrails, so results in the gated product may differ.
• The 0.0% prompt injection result covers only the evaluated attack set in one benchmark.
• DeepMind says it is still iterating on guardrails and is taking part in the U.S. government's voluntary pre-release testing process. No test results are given.
• The price is called "introductory" with no end date.