Gemini 4 Neon Is Now Available in Vercel's AI Gateway for Coding Agents

- Reading
- 5 min
- Code
- 1
Vercel's AI Gateway now lists Gemini 4 Neon, a faster version of Google's model built for code, agents, and multi-step tasks. Here's how to swap the model ID in your project without rewriting the integration.
Vercel updated the AI Gateway changelog to include Google's Gemini 4 Neon as another available model ID in the service. The company describes the model as a leap over the lighter Gemini 4 that was already in the catalog, promising output quality close to the larger models in the Gemini 4 family, but with faster responses. The stated focus is code, agent workflows, and tasks that require multiple sequential steps.
For front-end developers who already use the AI Gateway as a single layer for calling models from different providers, the news isn't really about Gemini 4 Neon itself: it's about not having to rewrite anything to try it out. Switching models becomes a matter of changing a string.
What the AI Gateway Solves (and Why the New Model Fits In)
The AI Gateway is Vercel's layer that exposes models from multiple providers (OpenAI, Anthropic, Google, among others) behind a single API compatible with the AI SDK. Instead of integrating each provider's SDK separately, each with its own authentication and response format, the app talks to the gateway and switches models by changing model.
This matters because switching generative AI models today is as much a product decision as a technical one: cost, latency, and quality change month to month, and the provider that was best yesterday might not be the best after the next update. Having Gemini 4 Neon as just another entry in the same catalog is the kind of update that doesn't require deploying new code, just an A/B test of the model ID.
Inside Gemini 4 Neon: 1M Token Context and Four Reasoning Levels
According to Vercel, Gemini 4 Neon accepts text, image, video, and documents as input, with a context window of 1 million tokens. This matters for anyone building agents that need to load an entire repository, a documentation base, or a long conversation history without truncating the prompt.
The model also supports:
- Tool calling: the model decides when to call an external function (fetch data, run code, query an API) instead of just generating text.
- Structured outputs: the response can be forced to follow a schema (JSON, for example), which avoids fragile free-text parsing on the front end.
- Grounding with Google Search and Google Maps: the model can anchor its response in real-time search or map data, reducing hallucination for factual or geographic questions.
- Four reasoning levels:
low,medium,high, andxhigh, which work as a dial between speed and reasoning depth per call.
This last point is what sets Neon apart from a merely "fast" model: the idea is to let whoever is calling the API decide, per request, whether they want an immediate response (low) or more costly, slower reasoning (xhigh) for a task that justifies the expense.
How to Integrate: The Model ID and an Example With AI SDK
The model identifier in the AI Gateway is google/gemini-4-neon. In the example Vercel itself publishes in the changelog, the integration via AI SDK looks like this:
import { streamText } from 'ai';
const result = streamText({
model: 'google/gemini-4-neon',
prompt: 'Write a TypeScript function that debounces async calls.',
});
for await (const text of result.textStream) {
process.stdout.write(text);
}The pattern is the same as any other model already configured in the AI Gateway: streamText() receives model and prompt, and the output is consumed as a stream, token by token, via result.textStream. There's no Google-specific SDK to install, nor a separate API key to manage: authentication and billing go through Vercel's gateway.
In practice, this means a project that already has streamText() pointing to another model, say a Claude or a GPT, can try out Gemini 4 Neon just by changing the model string value. For anyone deciding between low, medium, high, and xhigh, it's worth treating this like any other configuration parameter: measure latency and output quality for the actual use case before locking in a level in production, rather than assuming xhigh is always the right choice.
Beyond the Browser: Codex, Cursor, and Claude Code
Vercel also mentions that Gemini 4 Neon can be used inside coding agents such as Codex, Cursor, or another supported agent, with a specific guide for that (the "coding agents guide" cited in the changelog). This places the model in two quite different usage contexts: as the engine behind an AI feature inside an app via AI SDK, and as the engine of the agent itself that the dev uses in the editor to generate and review code.
For front-end teams that already have a coding agent set up, the practical question is whether Gemini 4 Neon is worth swapping in as the agent's default model, given Vercel's stated positioning of quality close to the larger models with faster responses. The changelog doesn't publish a comparative benchmark, so this is a decision that can only be settled by testing it in the actual workflow, ideally in the model playground that Vercel itself provides for trying out the model before committing to an integration.
What's Still Unclear
The material published by Vercel doesn't include per-token pricing, doesn't compare numerical latency between Gemini 4 Neon and the larger models in the family, and doesn't detail specific rate limits for the new model ID. Before migrating a production feature to google/gemini-4-neon, it's worth checking these points directly in the AI Gateway dashboard, since cost and usage limits vary by provider and by the chosen reasoning level.
There's also no indication so far that Google Maps grounding support is available for all regions or plans of the AI Gateway. Anyone relying on this feature for geolocated front-end features should confirm availability before designing the UI around it.
Translated from the Brazilian Portuguese original · Read the original
Vercel launches eve 1.0, AI agent framework with native React and Next.js integration
After months in public testing, eve reaches version 1.0 with native integrations for Next.js, SvelteKit, and other frameworks, promising response turns up to 7.3 times faster and significantly lower infrastructure cost.





