Google expands Gemini 4 Carbon for complex reasoning and workflows

Expanding access to Carbon

Google has officially expanded access to its Gemini 4 Carbon artificial intelligence model. This release targets a wide audience including developers, businesses, scientists, and general consumers. The company intends for the model to handle tasks that require multiple steps and extended reasoning capabilities.

While many AI models excel at quick responses, Carbon is built for long-running workflows. It is specifically designed to manage knowledge work, software engineering, and scientific research. These are areas where the AI must maintain consistency over a long period of time.

The model is now available through the Gemini API in Google AI Studio. It is also accessible to paying enterprises and developers through Gemini for business. This rollout follows an earlier limited release to specific cybersecurity defenders through Google's Fairwind programme.

Reasoning and verification

A core differentiator for Carbon is its ability to perform repeated verification. This makes it suitable for complex coding and debugging, where a single error can break an entire system. By checking its own work, the model aims to provide higher reliability for technical users.

Google DeepMind describes Carbon as its strongest model for cybersecurity and scientific research. It is meant to handle open-ended research questions that require deep logical leaps. The model is built to navigate these complexities without losing the thread of the original goal.

In terms of cost, Google has set the introductory price to match Gemini 4 Argon. It costs $2 per million input tokens and $10 per million output tokens. While Argon remains the standard for most knowledge work, Carbon is positioned for more demanding and intensive tasks.

Benchmark performance

Google reports that Carbon has achieved top scores across several industry benchmarks. It ranked first on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench. These tests measure how well the model handles coding, information retrieval, and automated workflows.

The company also highlighted the model's success in financial research. During testing before its official launch, the model performed as well as the leading option on Hebbia's financial research benchmark. It also topped 23 different model configurations on the Balyasny Asset Management financial intelligence benchmark.

In the realm of security, Google's Project Fortify system used the model to find all threats during vulnerability scanning. It also achieved a 95 per cent accuracy rate when sorting through those issues. These figures suggest the model is highly effective at identifying and sorting through security threats.

The swarm feature

For developers working on massive projects, Google is previewing a new '/swarm' feature in Google Antigravity. This functionality allows multiple AI agents to work on different parts of a single task at the same time. It moves the AI from a single-user interaction to a collaborative system.

In this setup, Carbon acts as the coordinator for the various agents. It can handle high-level roles such as system design while other agents perform specific sub-tasks. This allows for a more modular approach to solving complex engineering problems.

This multi-agent coordination is intended to streamline large-scale software development. By breaking a project into smaller pieces, the system can process more information simultaneously. This architecture helps manage the complexity of modern software architecture and data processing.

Rollout and availability

The wider rollout of Carbon is happening across several Google platforms. Users with AI Ultra subscriptions can access it through Google Antigravity, the Gemini app, and AI Mode in Search. This ensures that high-end users have immediate access to the new reasoning capabilities.

For businesses, the integration into Gemini for business provides a path for enterprise-grade deployment. This allows companies to incorporate Carbon into their existing workflows for research and development. It serves as a bridge between experimental AI and production-ready tools.

As Google continues to scale Carbon, the focus remains on its ability to handle demanding tasks. By providing a clear pricing structure and diverse access points, Google aims to make high-reasoning AI a standard tool for professional environments. The expansion marks a significant step in making complex AI workflows more accessible.