Vercel 携手 OpenAI 发布 GPT-5

The GPT-5 family of models released today, are now available through AI Gateway and are in production on v0.dev. Thanks to OpenAI, Vercel has been testing these models for a few weeks in v0, Next.js, AI SDK, and Vercel Sandbox.

From our testing, GPT-5 is noticeably better at frontend design than previous models. It generates polished, balanced UIs with clean, composable code. Internally, we’ve already begun using GPT-5 for Vercel's in-dashboard Agent and v0.dev/gpt-5. GPT-5 stands out in several ways for building agentic applications. We have been impressed with the long-context reasoning, and parallel tool usage capabilities which were useful for Vercel Agent.

How we're using GPT-5 in production with v0

We evaluated GPT-5 inside v0's composite model architecture, and it consistently impressed. Code quality matches or exceeds the best models available, and its frontend sensibilities stand out. It creates well-structured UIs from short prompts, often better than older models even with prompt tuning.

The smaller GPT-5 variants, mini and nano, also excel in low-latency delivery and their consistency. They’re already in use for our speed-sensitive codegen workloads, and perform well across tool calling and generation tasks.

To show what’s possible with GPT-5 on our AI infrastructure, we’ve built two interactive environments you can immediately get your hands on to put it to the test.

An open-source code-generation platform, for building real apps

Today we’re open-sourcing a vibe coding agent built on the Vercel AI Cloud. We’re essentially open-sourcing a v0-like app that you can clone, customize, and build on.

Under the hood, and abstracted away behind easy to use SDKs and a few lines of code, it's powered by Vercel's AI infrastructure and GPT-5 and other models for code generation, designed to be cloned, forked, and extended, and overall demonstrates the complete workflow, from submitting an initial prompt to watching responses stream in real-time as the agent generates code.

This demonstrates the end-to-end AI workflow as the reasoning model can not only determine what code to generate, but generate the files, install dependencies, run shell commands, and return a live, fully secure sandbox environment.

The platform's architecture is designed to support prompt handling, with reasoning powered by Fluid compute when running on Vercel Functions, through to untrusted code generation and execution in production.

Building an open-source vibe coding platform with GPT-5 and Vercel AI Cloud

Vibe coding platform's frontend is built primarily using Next.js and AI SDK, which to set up is just a few lines of code. When a user submits their message, it's sent via API sendMessage is triggered with both the prompt and the selected model, for which the platform defaults to GPT-5, are sent to the API endpoint. When code is deployed to Vercel, Framework-defined infrastructure determines the appropriate optimal resources to deliver the frontend quickly, via Vercel CDN, through to handling the API side of the calls.

The API is simply a function, that will run in Vercel Functions with Fluid compute which is optimized for prompting. AI calls commonly take a lot of thinking time, when the AI is reasoning. With Fluid compute, that idle time is either used for another function call (somebody else making a request, with their own prompt), or optimized to not charge CPU rates if not in use at all.

const result = streamText({
  model: "openai/gpt-5",
  system: prompt,
  messages,
  stopWhen: stepCountIs(20),
  tools
});

The API call, running AI SDK is as simple as creating a POST endpoint, receiving the incoming messages (all messages in the chat so far), confirming the user is not a bot with Vercel BotID. When deployed on Vercel, by using a string for the model we automatically leverage AI Gateway, to removing the need to manage multiple API keys.

As the model begins to return the generated code, it's configured in a newly spawned the sandbox that's a fast, stateless, ephemeral environment with no access to your projects or sensitive data, enabling it to safely run commands and arbitrary code that expire after a short timeout without lingering or unintended side effects.

As the sandbox runs, orchestrated by the AI agent, the sandbox can stream back data to the frontend. For example, as the backend is executing commands on the sandbox, the frontend can be informed to relay the latest activity to the user in real time:

{
  type: "data-run-command",
  data: {
    status: "done",
    sandboxId: "sbx_123",
    command: "npm install",
    commandId: "cmd_abc",
  }
}

That way, the UI updates in real time, without waiting for the entire task chain to complete.

Finally, it's important with potentially high-value API calls to mitigate abuse with sophisticated bot protection powered by BotID and rate limiting with Vercel Firewall.

Get started with GPT-5 via AI Gateway directly

import { streamText } from 'ai'

const result = streamText({
  model: "openai/gpt-5",
  prompt: "why is the sky blue?" 
})

The platform lets you compare three GPT-5 models (main, mini and nano), along with models from other providers for a head to head comparison.

Use v0.dev, or build your own

v0.dev now offers GPT-5 as an available model for improved UI generation, but this platform shows you how to build your own version from scratch. Whether you're scaffolding backend APIs, building infra-as-code, or generating fullstack starter kits, this system gives you a pattern to follow.

All components, AI Gateway, AI SDK, Sandbox, and BotID, are part of the Vercel AI Cloud. They are designed to be used together or dropped into existing projects. Everything runs from the Edge and can be integrated into existing applications.

Get Started

Clone the vibe coding platform, change the model, swap in your own tools, or connect it to a database. The infrastructure is in place to support it.

• Explore the code on GitHub

• Learn more about AI Gateway

Vercel AI Cloud is designed for developers shaping how software is created; building with models, not just using them.