Gemini 4 Carbon announced: Google’s long-running engineering model

Google DeepMind announced Gemini 4 Carbon as a new frontier model on 28 October 2026. The company says it designed the model for long-running engineering work and cybersecurity defense. Access begins this week with a limited group of cybersecurity defenders. Google DeepMind plans to gather feedback and improve safeguards before making the model available to developers, businesses and consumers.

Long-running tasks and coding take center stage

According to Google DeepMind, Carbon was developed to sustain complex, multi-step engineering tasks. The company’s own teams report that the model can work for days, even through dead ends, on tasks such as finding and fixing the root causes of problems in large codebases and low-level systems. This is reported as the experience of the company’s own teams.

In Google DeepMind’s announcement, the model ranks first in the recently updated v5.0 version of the Artificial Analysis Intelligence Index, with a score of 59. It is also said to have reached 1769 points and fourth place on WebDev Arena, with an increase of 89 Elo points. The company reports improvements in coding and agent-based tasks, as well as scientific work, full-stack web development and multimodal understanding.

Cybersecurity results and safeguards

Google DeepMind says Carbon delivers strong results across different stages of cybersecurity defense. In tests shared by the company, the model scored 88.2% on SEC-bench Pro, which focuses on finding vulnerabilities; 54.8% on ExploitGym, which tests whether vulnerabilities can be exploited; and 71.0% on CWE-bench, which focuses on generating fixes. According to the announcement, these benchmarks were evaluated without security restrictions in place.

In the company’s own tests, Carbon more than doubled its success rate at generating secure code compared with Argon. Google DeepMind also reports that successful replication attempts fell to 0% on Terminal-Bench 4.0, while attempts remained at 0.4%. The reported results also include a 0% attack success rate for the model in indirect prompt-injection testing. These are figures shared by the developer; on their own, they do not show performance under different usage conditions.

Access plans and announced pricing

In the first phase, access will be granted to selected cybersecurity defenders through the Fairwind Program. Google DeepMind will continue gathering feedback and improving safeguards before expanding availability. No access date has been announced for developers, businesses or consumers.

During the introductory period, pricing was announced at $2 per million input tokens and $10 per million output tokens. Cached input tokens will cost 95% less than standard input tokens. Once the introductory period ends, pricing is planned to rise to $4 per million input tokens and $20 per million output tokens. No end date has been specified for the introductory period.

Potential impact on architecture and visualization workflows

The announcement focuses primarily on engineering and cybersecurity use cases; it does not describe a specific scenario for architecture, interior design or visualization workflows. As a result, it cannot be assumed that Carbon directly builds 3D scenes, renders images or works in BIM software. Its focus on long-running coding and agent-based tasks may interest teams developing their own tools or technical automations, but the limited initial access will affect opportunities to try it.

For offices in Turkey, key points to monitor include when access will expand, how token usage will affect project costs, and compatibility with the software they use. The announcement provides no information about local hardware requirements or dedicated support for architecture software; pricing is based on token usage.