Google launches Gemini 4 Carbon with aggressive introductory pricing

A new entry in the market

Google officially released Gemini 4 Carbon on November 10, 2026. The model is now available across various platforms, including Google AI Studio for developers and Gemini for Business for enterprise users. It is also accessible to consumers who subscribe to Google AI Ultra.

The launch positions Carbon as a competitive alternative to existing models from Anthropic. Specifically, it aims to undercut the pricing of Claude Fable 5.2 and Claude Opus 5.5. By offering a lower price point, Google is attempting to capture developers who are sensitive to the costs of high-volume AI usage.

While the model is now public, it is not available for free users. Google has not yet announced whether it will eventually move to lower-priced tiers or offer downloadable weights for self-hosting. For now, it remains a premium offering for paying users and organizations.

The introductory price gap

The most striking feature of the launch is the temporary pricing structure. Until November 30, Carbon costs $2 per million input tokens and $10 per million output tokens. This makes it roughly one-quarter of the price of Claude Fable 5.2, which currently costs $8 per million for inputs and $40 per million for outputs.

This pricing advantage is scheduled to end on December 1. At that point, Carbon's rates will double to $4 per million input tokens and $20 per million output tokens. This shift will put Carbon at the exact same price point as Claude Opus 5.5, narrowing the price gap with Fable 5.2 from fourfold to twofold.

As Yahoo Tech reported, these differences are significant for high-volume users. Because Google bills for thinking tokens—the internal reasoning steps an AI takes before providing an answer—the lower per-token cost can result in substantial savings for complex workflows. However, these savings are only guaranteed during the current introductory window.

Mixed performance benchmarks

The performance of Carbon compared to Claude Fable 5.2 is not uniform across all tasks. In tests for latency-sensitive workloads, Fable 5.2 is faster, completing tasks in 16.9 minutes compared to Carbon's 26.1 minutes. Despite being slower, Carbon is cheaper for these high-effort tasks, costing an estimated $7.76 per task against Fable's $9.08.

The models also show different strengths depending on the specific type of work being performed. Carbon leads on terminal benchmarks, scoring 65.2% on the Terminal-Bench 4.0 compared to Fable's 59.6%. This suggests it may be more effective for automation tasks that occur within a command-line interface.

Conversely, Fable 5.2 holds the advantage for repository-level coding tasks. On the FrontierCode-AA 2.0 benchmark, Fable scored 48.8% while Carbon scored 40.6%. These results indicate that the best model depends entirely on whether a user needs to automate simple commands or manage complex codebases.

Calculating total workflow costs

When evaluating these models, the published token price is only one part of the equation. Users must also consider the cost of retries, human review time, and the number of reasoning steps required to reach a successful output. A model that appears cheaper on paper might become more expensive if it requires more human intervention to correct errors.

Google's documentation confirms that billed output includes the tokens used during the model's thinking process. This means that for highly complex reasoning tasks, the output costs can scale quickly. Teams need to account for these hidden variables when comparing Carbon to its competitors to ensure they meet their specific service level agreements.

Because the pricing changes on December 1, businesses are encouraged to run pilots immediately. A controlled pilot using identical tasks and permissions will provide more accurate data than any marketing comparison. This allows teams to see how Carbon handles their specific workflows before the introductory discounts expire.