Google DeepMind unveils Gemini 4 Carbon, led by cybersecurity gains

Cybersecurity results at the center

Google DeepMind says Carbon shows significant gains in coding and agentic tasks, as well as in capabilities critical for cybersecurity defense. The company reports a firm lead on three external benchmarks, evaluated without guardrails:

• SEC-bench Pro: 88.2% for finding vulnerabilities
• ExploitGym: 54.8% for exploitation
• CWE-bench: 71.0% for patching, on Collinear's benchmark

On internal benchmarks, Carbon achieves more than 2x uplift in success rate for generating secure code over Gemini 4 Argon, according to Google DeepMind. The company also says the Chrome team reported that Carbon performed more than 10X better than another lab's large frontier model when generating high quality patches. Wiz, Google's cloud security platform, observes Carbon leading on benchmarks for web penetration testing, which the company describes as a core capability for Wiz's Scan for Good initiative.

Built for long-horizon engineering work

Google DeepMind describes Carbon as designed for much harder, long-horizon engineering problems and sustained agentic workflows. Internally, its teams have seen the model run for days and persist through dead ends on complex, multi-step tasks such as root-causing and patching subtle bugs across large codebases and low-level systems.

The company says it focused on coding, agentic and scientific tasks, as well as full-stack web development and multimodal understanding. In web development, Carbon gained +89 elo points to land at #4 (1769) on WebDev Arena.

Safety and alignment claims

Carbon uses the same multi-layered safety and security stack as Argon, which combines model training, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring. Google DeepMind reports several alignment results:

• Successful cheating on Terminal-Bench 4.0 fell to 0.0%, with cheating attempts at 0.4%.
• High-severity misalignment alerts in live coding environments with internal teams dropped by almost 80%.
• Unsupported claims were cut by approximately 2x compared with Argon in internal testing.
• Overreach fell by 26% and under-service by 29%.
• On Gray Swan's Indirect Prompt Injection benchmark, Carbon reached a 0.0% attack success rate, versus 0.7% for Argon.

The company adds that, when facing ambiguity, Carbon retrieves relevant sources and surfaces discrepancies to the user before acting.

Limited rollout and pricing

Access is being phased. Starting this week, Carbon is available to a subset of cyber defenders, with additional trusted partners added gradually through the Fairwind Program. Google DeepMind says it is engaged in the U.S. government's voluntary process for pre-release model testing and is gathering feedback and iterating on guardrails before opening Carbon to developers, businesses and consumers through its product surfaces "as soon as possible".

Carbon will launch at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off the input token price. Argon launched two weeks before this announcement, according to the company.

What it means

For now, the model is reserved for security professionals, so most developers and companies cannot use it yet. The company's focus on vulnerability detection, validation and remediation signals where it expects early value: defenders working across the full cybersecurity lifecycle. Broader availability depends on the outcome of feedback and guardrail work that Google DeepMind says is still under way. All benchmark figures cited here come from the company's own announcement.