SIAgentNews.com
← All news
DevelopingGeminiAgents

Google launches Gemini 4 Carbon for complex autonomous AI agents

The new model is designed to handle long-horizon tasks by coordinating multiple specialized AI agents to plan, build, and verify work.

· First reported by Brave New Coin

one-sixthcost of the highest-scoring model on the MKA benchmark

The short version

  • Gemini 4 Carbon is built for 'long-horizon' tasks requiring multiple stages of planning, execution, and verification.
  • The /swarm feature in Google Antigravity allows a single prompt to trigger a team of specialized agents working in parallel.
  • The model achieved more than twice the cost efficiency per task compared to the next-best external model on key benchmarks.
  • Carbon supports multimodal inputs and features a context window of up to one million tokens.
  • Pricing for API customers is set at $2 per million input tokens and $10 per million output tokens through November 2026.

Why it matters

This shift from single-prompt responses to multi-agent coordination means AI can now tackle projects that require sustained reasoning over hours rather than seconds. It allows for more reliable results in high-stakes fields like cybersecurity and software engineering by incorporating automated verification steps.

A shift toward long-horizon tasks

Google has initiated a wider rollout of Gemini 4 Carbon, a frontier AI model specifically engineered for complex agentic workflows. Unlike standard AI assistants that focus on responding to individual prompts, Carbon is built for long-horizon tasks. These are projects that require a sequence of planning, execution, testing, and refinement to reach a final goal.

The model builds on the Gemini 4 Argon architecture but utilizes more reasoning tokens to investigate problems deeply. This extra computational effort allows the system to evaluate various solutions and perform repeated verification rounds. According to Google DeepMind, this makes the model better suited for broad scientific inquiries and technical engineering jobs that need to interact with various software tools.

While Carbon is the most powerful option for demanding assignments, Google notes that Argon remains a viable choice for standard knowledge work. Carbon is intended for scenarios where the need for deeper reasoning justifies the higher computational requirements. It supports multimodal inputs, meaning it can process text, images, audio, and video simultaneously.

Coordinating teams of AI agents

A major component of this release is the /swarm feature within Google Antigravity, an agent-focused development platform. This tool allows a user to convert one request into a group of diverse AI agents. These agents can divide work, execute tasks in parallel, and review each other's output to ensure accuracy.

The system employs a coordinating agent to distribute responsibilities among parallel workers that operate in isolated Git workspaces. Additional verification agents are also included to run tests and identify potential problems before any work is finalized. Google describes this approach as a system where agents plan, build, and verify in parallel to complete a project.

Developers can monitor this team's progress through a shared message board and a dedicated interface. This setup allows for human intervention if necessary while the agents handle the heavy lifting. Google demonstrated this by creating a 3D exhibition of the Eiffel Tower's construction using historical photographs and engineering drawings as references.

Performance in high-stakes fields

Google reports that Gemini 4 Carbon ranks first on several key benchmarks, including CursorBench 4.0 and Zapier’s AutomationBench. The company claims the model is over twice as economical per task as the next-best external model across these evaluations. These tests specifically measure agentic coding, cybersecurity, and complex knowledge work.

In the Multimodal Knowledge Agent (MKA) benchmark, the model ranked second in overall quality. This evaluation involved 200 tasks designed to reflect work that would typically take a human professional between three and eight hours to complete. The results highlight Carbon's ability to balance high-quality output with the computational costs of complex workflows.

Cybersecurity is another primary focus for the new model, with specific scores reported for vulnerability discovery and patching. Google initially restricted these specialized capabilities to verified security defenders through its Fairwind Program. This strategy aims to support defensive research while maintaining safeguards against the potential misuse of the model's capabilities.

Availability and pricing structure

Gemini 4 Carbon became available to a broader audience on November 10, following a period of limited rollout. Users can find the model within the paid Gemini API options in Google AI Studio and Android Studio. Enterprise customers can utilize it through Gemini for business, while Ultra subscribers can access it via the Gemini app.

For API customers, Google has set an initial price of $2 for every million input tokens and $10 for every million output tokens. These specific rates will remain in effect through November 30, 2026. After that date, the price is scheduled to rise to $4 for every million input tokens and $20 for every million output tokens.

Google is also offering a special opportunity for the scientific research community. Scientists who qualify can request access to Carbon and Argon via Antigravity’s top-tier usage plan for a six-month period at no cost. This program is subject to selection, capacity, and other specific program requirements as part of Google's broader expansion into scientific research tools.

What to watch

  • How the /swarm feature handles conflicts when multiple agents propose different solutions.
  • The effectiveness of Carbon's automated security testing in preventing sophisticated prompt-injection attacks.

Try it

If you are a developer, check the Google Antigravity platform to see if you can access the /swarm preview for multi-agent coordination.

Sources and links

First reported by Brave New Coin. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news