Google DeepMind has announced its next frontier model: Gemini 4 Carbon. Described as the company's strongest model to date, it features state-of-the-art cybersecurity capabilities and excels at long-horizon tasks.
Carbon ranks #1 on the recently updated v5.0 of the Artificial Analysis Intelligence Index with a score of 59. It demonstrates significant gains across coding, agentic tasks, and key cybersecurity defense capabilities. In Google's internal benchmarks, Carbon drives a strong uplift in generating secure code, while live coding environments showed reduced misalignment and fabrication rates.
Building on Gemini 4 Argon, which was launched recently, Carbon illustrates the pace at which Google is pushing the frontier forward. The company is taking a deliberate approach to its rollout, actively engaging in the U.S. government’s voluntary process for pre-release model testing. Access will be gradually phased to trusted partners via the Fairwind Program before a broader release to developers and businesses. Gemini 4 Carbon will launch at $2 per million input tokens and $10 per million output tokens (with cached input tokens priced at 95% off).
Improving coding and reasoning for challenging, long-horizon tasks
Designed to take on much harder, long-horizon engineering problems, Carbon is built for sustained agentic workflows. Google's internal teams have observed Carbon running for days, persisting through dead ends to tackle complex, multi-step engineering tasks such as root-causing and patching subtle bugs across large codebases and low-level systems.
With improved output quality in web development, Carbon also jumped +89 Elo points to land at #4 (1768) on the WebDev Arena.
State-of-the-art cybersecurity performance
Gemini 4 Carbon delivers a significant leap in cybersecurity capabilities, achieving state-of-the-art performance across the defense lifecycle: vulnerability detection, validation, and remediation. The Chrome team reported that Carbon performed more than 10X better than other large frontier models when generating high-quality patches, and cloud security platform Wiz has also observed Carbon leading in web penetration testing.
Benchmark / Evaluation | Carbon Score | Domain / Focus |
|---|---|---|
Artificial Analysis Index (v5.0) | 59 (#1 Rank) | General AI Intelligence |
SEC-bench Pro | 88.2% | Finding real vulnerabilities |
ExploitGym | 54.8% | Exploitation & Risk Validation |
CWE-bench | 71.0% | Root cause patching |
WebDev Arena | #4 Rank (1768 Elo) | Web Development |
Gray Swan’s IPI | 0.0% attack success | Prompt injection resistance |
Advancing frontier model safety & security
Gemini 4 Carbon builds on the multi-layered safety stack of Argon, combining model training, real-time misuse defenses, and prompt-injection resilience. Google challenged itself to go a step further on alignment, training the model to respond to complex tasks correctly:
Reducing shortcuts: Carbon mitigated shortcut attempts in training, reducing successful cheating on Terminal-Bench 4.0 to 0.0% while improving overall accuracy. High-severity misalignment alerts dropped by almost 80% in live coding environments.
Greater transparency: Internal testing showed Carbon cuts unsupported claims by approximately 2x, putting methodological caveats directly into the deliverable rather than fabricating missing evidence.
Improved collaboration: Instead of guessing ambiguously, Carbon retrieves relevant sources and surfaces discrepancies to the user before acting.
Resisting prompt injection: Carbon successfully resisted every evaluated attack attempt on the Gray Swan Indirect Prompt Injection Benchmark.
Google's goal is to deliver meaningful leaps in software engineering and cybersecurity defense without compromising on safety, rolling out to developers and consumers in the near future.
