Google launches Gemini 4 Neon for analysis and coding, with access tied to subscription plans
The new model is rolling out in Gemini, AI Studio and Antigravity. Standard API pricing is $3 per million input tokens and $15 per million output tokens, with capabilities and access varying by product.

Google began rolling out Gemini 4 Neon on November 18, introducing an artificial intelligence model aimed at document analysis, software development and work involving multiple steps. The launch expands the choices for Google AI Pro and Ultra subscribers, as well as businesses and developers building Gemini into their own services.
The company positions Neon between Hydrogen, intended for fast everyday tasks, and Carbon, designed for more demanding, extended projects. Its pitch is advanced reasoning with shorter execution times, although the cost, available capacity and access arrangements depend on the product being used.
The release notes for the API, the interface that connects the model to other applications, list it as generally available from Wednesday. For the consumer Gemini app, Google announced a rollout to Pro and Ultra plans. Access to the new model varies across accounts and products.
Who can use Neon
In the Gemini app, Google AI Pro and Ultra subscribers can choose Neon from the model menu, according to the company's announcement. Google is promoting it for writing, file analysis and research, including assignments that draw on information from several documents.
Developers can use it through the paid Gemini API. AI Studio's documentation also provides for access through Google AI Pro and Ultra plans. Businesses can obtain the model through Gemini for business, using either its API or the relevant model selector.
Antigravity, Google's environment for working with coding agents, has its own eligibility rules. Its documentation includes Neon with Ultra, paid Pro subscriptions and selected trials linked to eligible premium devices. Free and Google AI Plus plans do not include that access.
The availability of a model therefore needs to be checked within the product being used. App subscriptions, usage-based API billing and promotional eligibility represent different routes to accessing Neon, with their own conditions.
Carbon's temporary discount changes the price comparison
Neon's standard API rates are $3 per million input tokens and $15 per million output tokens. Tokens are the units into which the system divides the content it processes and generates. A word can occupy one token or several.
Google's pricing table includes reasoning tokens in the output charge. It lists a rate of $0.15 per million input tokens reused through implicit caching. That rate applies to content recognized by the caching mechanism; the model does not support explicitly creating a cache.
The comparison with Carbon depends on the date. Google's current table lists promotional Carbon rates of $2 per million input tokens and $10 per million output tokens through November 30, 2026. It lists rates of $4 and $20, respectively, from December 1.
Neon's published standard token prices are therefore below the ordinary Carbon rates scheduled for December but above Carbon's current promotional prices. The final bill for a task also depends on the number of tokens consumed and the tools used. The API has separate charges for certain search queries.
Those rates apply to API use. Subscriptions providing access to the apps have their own terms; a price per million tokens is not the monthly subscription price for Google AI Pro or Ultra.
What the million-token context window covers
The API's model specification lists a combined input and output context window of 1,048,576 tokens. That is the space available for material the model receives and produces within the interaction. Supported inputs include text, images, video, audio and PDFs; the specified output is text.
The default synchronous output limit is 65,536 tokens. Google also offers a long-decode mode that can reach the combined context-window limit. The advertised overall capacity and the size of a response under the default mode are separate limits.
The launch coincides with another API update. From November 18, it accepts Microsoft Office and OpenDocument files and allows users to select particular parts of documents, spreadsheets, videos and audio. The release notes list pages, worksheet tabs, cell ranges and time intervals among the available options.
For people working with lengthy files, selecting relevant sections allows them to send those parts rather than the entire contents. That update is announced for the API; the features available in an individual application depend on its integration.
Antigravity divides planning and execution between models
In Antigravity, Neon can work on its own or as part of Orchestrator (Neon). The latter combines two models: Neon handles planning and review, while Hydrogen makes file changes, writes code and runs checks.
The documentation describes control passing between the models as work progresses. If Hydrogen encounters a problem requiring a revised plan, it hands the task back to Neon. Coordination takes place within the same conversation, without the user having to switch models manually for each stage.
Antigravity gives Neon a context window of up to 512K tokens, below the capacity published for the API. When selected directly, Neon offers four reasoning levels: Low, Medium, High and Extra. High is the default, while Extra corresponds to the xhigh setting in the command-line interface.
Orchestrator uses a fixed configuration and does not offer that reasoning-level choice. In new conversations on eligible plans, Antigravity defaults to Neon if the user has not chosen another model. It preserves an existing selection.
Evaluations show gains alongside continuing limitations
Artificial Analysis published an evaluation on November 18 giving Neon at xhigh reasoning a score of 61 on its Intelligence Index. It placed third in that comparison, one point above Carbon at high reasoning. Those results apply to the tests and configurations identified by the evaluator.
The evaluator also calculated a lower cost per task than Carbon when using Carbon's ordinary, undiscounted prices. The current promotion changes the price comparison facing an API customer today.
Google DeepMind's model card warns that Neon can produce false or fabricated information. It gives a knowledge cutoff of March 2026 and says knowledge in some fields may be less current. Access to searches and recent documents is relevant when answering questions about subsequent events.
The card says a full safety and security report will be released in the coming days. The launch gives users another option for processing information and developing software. Results will depend on the task, the configuration and review of the answers and code produced.