Google DeepMind launches Gemini 4 Carbon for cybersecurity

A new frontier model

Google DeepMind recently announced Gemini 4 Carbon, a new frontier model that arrives just two weeks after the release of Argon. This model is positioned as a state-of-the-art solution for long-horizon tasks and cybersecurity applications. It represents a significant step forward in the Gemini family's capabilities.

The model has achieved a top ranking on the Artificial Analysis consolidated benchmark. This puts Google DeepMind in the lead for the first time in the history of that specific independent leaderboard. It outperformed several other major models across various intelligence and coding indices.

Technical specifications for Carbon include a 1M-token context window and a 1M maximum output limit. It is designed to be multimodal, accepting inputs in the form of text, images, video, and speech. While it handles these diverse inputs, its output remains strictly text-based.

Cybersecurity performance

A major focus of the Carbon launch is its performance in security-related tasks. According to the announcement, the model achieved an 88.2% score on SEC-bench Pro, which measures the ability to find vulnerabilities. It also scored 71.0% on CWE-bench for patching and 54.8% on ExploitGym for exploitation tasks.

Google claims that the model shows improved safety and reliability compared to its predecessor, Argon. Specifically, unsupported claims fell by about 2x, and overreach decreased by 26%. These metrics suggest a focus on reducing the frequency of the model providing incorrect or out-of-scope information.

Independent evaluations also highlight the model's strengths in specific domains. It currently holds the #1 spot on the Finance Index and the Legal Research Benchmark. It also leads the Vending-Bench 2 and ties the human baseline for 3D spatial understanding on the Blueprint-Bench 2.

Deployment and pricing

Access to Gemini 4 Carbon is being rolled out in a phased approach. It is first being made available to trusted cyber defenders through the Fairwind Program. Google has promised that wider developer access will be provided as soon as possible following this initial launch.

Pricing for the model is expected to be competitive for enterprise users. Some reports suggest a standard pricing of $4 per million input tokens and $20 per million output tokens. However, there is a 50% launch discount currently in place with no specified end date.

The model also offers a cached input cost of $0.20 per million tokens. This pricing structure is intended to help developers manage costs while utilizing the model's large context window for complex tasks. These costs are lower than some other leading models on the market for similar index tasks.

Benchmarks and limitations

While Carbon leads in many areas, it faces competition in specific coding benchmarks. On the FrontierCode v1.1 index, it scored 51% compared to 56% for Opus 5.5. It also trails GPT-6 Astra on the FrontierSWE benchmark, scoring 61.2%.

The model's hallucination rate is a point of discussion among early users. While it leads the Omniscience index with a score of 54, its hallucination rate rose from 4% on Argon to 11% on Carbon. Despite this increase, it remains the second lowest measured rate among the top models.

Some users have expressed concerns regarding the risk of hallucinations in cybersecurity contexts. Additionally, some analysts suggest that Google may still be a few months behind the next major releases from competitors like OpenAI and Anthropic. These factors suggest that while Carbon is a leader, the competitive landscape remains dense.

The future of Gemini

The launch of Carbon follows a trend of increasing training and launch velocity at Google. CEO Sundar Pichai noted that AI development is accelerating due to internal research and infrastructure improvements. This suggests that the pace of new model releases will likely continue to quicken over the next year.

Internal workflows at Google are also evolving to handle these rapid launches. Engineers describe using overnight agent loops and swarms to inspect reinforcement learning data. In this environment, new model launches are beginning to feel like periodic checkpoints in a continuous development cycle.

For the cybersecurity community, the integration of Carbon into the Fairwind Program is the immediate next step. This allows for real-world testing by defenders before the model reaches the general public. This staged rollout ensures that the model's high-stakes capabilities are vetted by professionals first.