Google Opens Gemini 4 Carbon to Paid-Tier Developers

News Summary

Google opened Gemini 4 Carbon to developers on November 10, 2026, positioning it as a model for extended, multi-step work in software engineering, full-stack web development, scientific research and cybersecurity. Access starts through the paid tier of the Gemini API in Google AI Studio and Android Studio; Google AI Ultra subscribers can also try it in AI Studio and the Google Antigravity agent-first development platform. Google has not described the release as unrestricted access across every product or account.

A model aimed at work that takes multiple steps

Carbon builds on Gemini 4 Argon and is designed for agentic workflows that can span a long sequence of actions. Google says the model can investigate and fix subtle bugs across large codebases, produce web applications, and work through open-ended research tasks. In its October launch announcement, the company described internal engineering runs that continued for days while pursuing difficult problems. Those examples are company-reported capabilities, not a guarantee that every task will complete successfully or without supervision.

For science and technology education, the longer workflow emphasis could make the model useful as a research and coding assistant: learners might ask it to explore a hypothesis, inspect code, or explain a sequence of debugging steps. Such use still calls for checking sources, methods and generated code. Google’s model card warns that Carbon can hallucinate, sometimes runs slowly or times out, and may consume more tokens at higher thinking settings.

Where paid developers can use it

The November 10 announcement names the paid Gemini API tier in Google AI Studio and Android Studio as developer entry points. Google AI Ultra subscribers are separately offered access in AI Studio and Antigravity. Its model documentation also lists Gemini for business and other distribution channels, while clarifying that availability depends on the product and subscription tier. Developers should therefore check the relevant product documentation and account access before planning a deployment.

Google announced introductory API pricing in October at $2 per million input tokens and $10 per million output tokens, with cached input tokens priced 95% below the input rate. Pricing and access terms may change, so teams should confirm current rates and quotas in the API console before estimating project costs.

What Google’s evaluations report

Google’s October announcement highlighted cybersecurity results: 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation tasks, and 71.0% on CWE-bench for patching. The company also reported that Carbon more than doubled Argon’s success rate for generating secure code on an internal benchmark. These are vendor-reported evaluation results; benchmark performance does not by itself establish how a model will perform in a particular codebase or operational environment.

The published model card reports a 1-million-token combined input-output context window and support for text, images, audio and video as inputs. It also includes evaluations in coding, computer use, knowledge work, science and math. For example, the card lists 80.0% on RiemannBench and 64.8% on Terminal-Bench-Science 0.1. The card says Carbon uses more thinking tokens than Argon for difficult reasoning, with configurable thinking levels intended to balance quality, latency and cost.

Why the release matters

The developer launch moves Carbon beyond its initial staged rollout to cyber defenders and trusted partners, and gives software teams a direct route to try its longer-horizon capabilities. Its strongest stated fit is work where a model must maintain context, take several connected steps and verify results along the way. In practice, teams will need to weigh that potential against token costs, latency, access limits and the familiar need for human review—especially when outputs affect security or scientific conclusions.