Gemini 4 Carbon: Our next frontier model

What it claims

Gemini 4 Carbon is Google DeepMind's frontier model, following Gemini 4 Argon by two weeks. DeepMind positions it for long-horizon agentic engineering and cybersecurity defense. Beyond the headline benchmarks, the release reports alignment and reliability gains: zero successful cheating on Terminal-Bench 4.0, zero prompt-injection success on Gray Swan's IPI benchmark, and a roughly 2x cut in unsupported claims over Argon. The model is not yet available to general developers.

Evidence

The three cybersecurity scores (SEC-bench Pro 88.2%, ExploitGym 54.8%, Collinear's CWE-bench 71.0%) were evaluated without guardrails. That is the setting, not the deployed product. No competitor scores are given for these benchmarks, only the claim of state of the art.

Other results stated on the page:

• WebDev Arena: +89 Elo to 1769, ranked #4. Carbon does not hold the top spot here.
• Secure code generation: more than 2x success rate over Argon, on internal benchmarks.
• Chrome team patches: more than 10x better than "another lab's large frontier model," which is not named. This is an anecdotal report.
• Wiz: observes Carbon leading on web penetration testing benchmarks. No figures are given.
• Terminal-Bench 4.0: successful cheating fell to 0.0% and cheating attempts to 0.4%, while accuracy rose. No accuracy figure is given.
• Gray Swan IPI Benchmark: 0.0% attack success rate, against Argon's 0.7%.
• Internal live coding: high-severity misalignment alerts dropped by almost 80%.
• Scope adherence: overreach fell 26% and under-service fell 29%. No baseline is named, presumably Argon.

The page says internal teams have run Carbon for days on multi-step tasks, such as root-causing bugs across large codebases and low-level systems. No metric accompanies this claim.

How it works

The page gives no architecture, size or training detail. It says Carbon uses the same safety stack as Argon: model training, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring. It adds training aimed at reducing shortcut-taking under pressure. Two behaviours are described. First, the model writes methodological caveats into its output, or lists blocked actions and what they still require. Second, it surfaces discrepancies to the user before acting on ambiguous tasks.

How to use it

Initial access goes to a subset of cyber defenders, then expands to "trusted partners" through the Fairwind Program. Developer, business and consumer access is promised "as soon as possible," with no date. Pricing is introductory and matches Argon: $2 per million input tokens and $10 per million output tokens. Cached input tokens are 95% off the input price, which comes to $0.10 per million. No API model name or endpoint is given.

Limits

Most alignment and coding figures come from internal benchmarks or partner reports without published methodology. The cyber benchmark scores were measured without the guardrails that will apply in deployment. DeepMind says it is still iterating on guardrails and taking part in the U.S. government's voluntary pre-release testing process, so behaviour at general release may differ. The price is labelled introductory, which implies it may change.