The News, Explained
Google introduced Gemini 4 Neon on November 18 and began its rollout. It designed the model for complex coding, analysis and multistep work, citing reports and summaries based on spreadsheets, PDFs and photos as examples. In the Gemini app, Neon is rolling out through the model menu for Google AI Pro and Ultra subscribers. Source
Personal-account help lists Neon for both Chats and Tasks under Pro and Ultra. In Gemini, Tasks is the mode for delegating multistep work. Auto routes prompts to Hydrogen or a lightweight model; advanced models require manual selection. Work and school availability depends on the Workspace license. Source
The app offers Low, Medium and High thinking levels to adjust reasoning depth. Manual selection defaults to Low unless a level is specified. Gemini 4 Neon access depends on the plan and remaining allowance; higher thinking levels and advanced models consume more usage. Source
Developers can use an API, a connection through which software sends requests to a model. Google’s announced rates are $3 per million input tokens, $15 per million output tokens and $0.15 per million cached input tokens. Tokens are units the model processes. Its context window—the information it can handle within a task—is one million tokens. Developers pay token-based API rates, while Neon app access follows the subscription and usage conditions. Source Source
Settings also qualify the cost claims. Google reports 32% lower cost per task than Carbon in benchmarks at Neon’s default high thinking setting, and introduces xhigh for harder work. These are Google’s evaluation results. App help separately lists Low, Medium and High with a manual-selection default of Low. Reading the announcement’s default high or xhigh as the app’s controls would mix different usage conditions. Source Source
OYOPICK’s Take
For existing Pro subscribers seeking deeper document analysis, Neon adds a welcome option. In a hypothetical comparison of items across several receipts, recording the app’s thinking level alongside the model name would make results easier to compare. The selected reasoning level and remaining usage form part of the working conditions. Source Source
We would assess API task cost and app subscribers’ choices against the way each reader works. Whether Google’s reported 32% reduction translates into savings for a particular job depends on the task and settings. We hope the ability to choose analysis depth develops alongside clearer, more predictable usage. Source Source