Web2View
Contact
IA aplicada,  Engenharia e operação

What to check in Gemini 4 Neon before connecting it to an agent

Author

Web2View

Published

Editorial illustration of an abstract portal that layers code, documents, images and video in reasoning stacks.

Vercel’s AI Gateway now offers Gemini 4 Neon to code agents and multi-step tasks. The published identifier is google/gemini-4-neon. The model ships with a 1-million-token window, accepts text, images, video and documents, supports tool calling and offers four reasoning tiers: low, medium, high and xhigh.

What Vercel published

We started from the 18 November 2026 changelog signed by Jerilyn Zheng and Zachary Chen. Vercel states that Gemini 4 Neon from Google is now on the AI Gateway. The text positions the model as an advance over the previous Gemini 4—lighter, with output quality close to the larger Gemini 4 models and faster responses. Intended use cases are programming, agent workflows and multi-step tasks.

At Web2View, we read the note through the lens of the contract it sets. Identifier, input formats and tools are what decide whether an agent can call Neon without a separate integration.

The changelog lists, for this identifier:

- 1-million-token context window;
- text, images, video and document inputs;
- tool calling and structured outputs;
- grounding with Google Search and Google Maps;
- reasoning tiers low, medium, high and xhigh.

AI SDK and Chat Completions

In the AI SDK example, streamText receives the model google/gemini-4-neon and a prompt. The response streams via textStream. In the sample, the identifier is passed directly in the model field.

Teams already using Chat Completions point the client to https://ai-gateway.vercel.sh/v1, authenticate with the AI_GATEWAY_API_KEY variable and send the same identifier in the model field. Vercel’s example also streams.

If your workflow already talks to the AI Gateway, the practical switch is just the identifier. We treat that as configuration, not a new project.

Claude Code, fx, Codex and Cursor

In the note, we also see the path for agents—not only for application code.

On fx, the tool is available on macOS and Linux for x86_64 and arm64. With no saved credentials, fx opens a selector: sign in with Vercel in the browser or paste an AI Gateway key. The /setup command reopens that choice. Then /model google/gemini-4-neon selects Neon. Signing in with Vercel surfaces the full AI Gateway catalog in the selector.

In Claude Code, the command npx vercel@latest ai-gateway setup detects installed agents, creates an AI Gateway key and writes the configuration. The setup enables model discovery from the gateway. After that, /model lists the catalog, and Gemini 4 Neon can be chosen there.

For Codex, Cursor or any other supported agent, the changelog does not describe each tool’s file. It points to the coding-agents guide.

What the DeepMind announcement adds

On the same day, 18 November 2026, Google DeepMind unveiled Gemini 4 Neon. Koray Kavukcuoglu, Senior Vice President at Google DeepMind and Google’s Chief AI Architect, describes the launch as frontier intelligence, with rapid completion of end-to-end tasks in complex software engineering, knowledge work and scientific workflows.

In the announcement, pricing is US$ 3 per million input tokens, US$ 0.15 per million cached input tokens—a 95 % discount—and US$ 15 per million output tokens. The announced window is also 1 million tokens.

Outside the gateway, Google says the model is available via the Gemini API in Google AI Studio, the Gemini for Enterprise API and model selector, to paid Google AI Pro and Ultra subscribers in Google Antigravity, and to Pro and Ultra subscribers in the Gemini app.

At the xhigh tier, Google states that Neon placed third on the Artificial Analysis Intelligence Index with 61 points, behind larger and slower models. We read these figures as positioning in the announcement, not as a guarantee of results in any specific repository.

What to separate when adopting

In practice, we separate three sources that blend easily:

- Vercel changelog: identifier, window, input types, tool calling, structured outputs, grounding, reasoning tiers and the AI SDK, fx and Claude Code paths.
- Coding-agents guide, merely referenced: Codex, Cursor and other supported agents.
- Google DeepMind announcement: pricing, Google ecosystem products and the Artificial Analysis index position.

For a team that already authenticates agents on the AI Gateway, the next step is to select google/gemini-4-neon and choose the reasoning tier. For teams that do not yet have a key or login, setup comes before the model. That is the order in which Vercel’s documentation structures fx and Claude Code.

If your team plans to connect a code agent or multi-model workflow to a product, talk to Web2View. Talk to Web2View

Latest articles

All articles →