GPT-5 arrives on Vercel's AI stack

OpenAI's GPT-5 family is now live on Vercel's AI Gateway and in production on v0.dev. Ahead of today's release, Vercel says it tested the models for several weeks across v0, Next.js, AI SDK, and Vercel Sandbox. In its evaluation, the company found GPT-5 notably stronger at frontend design than prior models, producing clean, composable code and balanced layouts from minimal prompts. The model also showed solid results inside agentic workflows, with long-context reasoning and parallel tool handling standing out while running Vercel's dashboard Agent.

Outside of GPT-5 itself, Vercel is highlighting the smaller variants: GPT-5 mini and nano. Both have proven effective for low-latency delivery and consistency, and are already handling speed-sensitive codegen workloads across tool calling and generation tasks.

A community playground and an open-source reference app

To let others evaluate GPT-5 without setup friction, Vercel has set up a dedicated model playground inside AI Gateway. It covers every Gateway model, including GPT-5, gpt-oss-20b, and gpt-oss-120b, and is free to use. From the same page, you can copy the model string and drop it into AI SDK, taking advantage of Gateway's rate limits, performance, reliability, and observability features.

Vercel also released an open-source "vibe coding" platform built on the Vercel AI Cloud. It functions as a v0-like app designed to generate real applications from a prompt to a live preview, demonstrating the complete flow: a prompt enters, code streams back, and a secure sandbox serves an instant preview. The app combines GPT-5 with AI SDK, AI Gateway, Vercel Functions, Sandbox, and BotID, and is available to clone and customize on GitHub.

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

How the vibe coding platform works

When a user submits a message, the sendMessage API is invoked. Both the prompt and the chosen model (GPT-5 by default) are sent to the backend. Once deployed to Vercel, the framework-defined infrastructure routes the frontend through the Vercel CDN and sends API requests to Vercel Functions.

The backend is a lightweight function running on Fluid compute, which suits prompt-based workloads where the model spends time reasoning. Fluid compute can reassign idle cycles to other requests or avoid charging for unused CPU during those gaps. The API call itself goes through AI SDK, receiving the full message history, verifying the user via Vercel BotID, and forwarding the request. By specifying the model as a string, the request routes automatically through AI Gateway, removing the need to manage separate API keys.

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

Code execution happens on a fresh, isolated sandbox that's stateless and expires after a short timeout. It has no access to your projects or data, making it safe to run arbitrary code. As the model streams responses back, the sandbox runs them and sends real-time updates to the frontend, so the UI reflects command progress instantly without waiting for the full task chain to finish. Abuse protection is handled by BotID and rate limiting via Vercel Firewall.

Starting with GPT-5

The open-source platform lets you switch between all three GPT-5 models (main, mini, and nano) alongside models from other providers for side-by-side comparison. v0.dev also now offers GPT-5 as a UI generation model option.

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

The pattern can extend beyond frontend generation: scaffolding backend APIs, generating infrastructure as code, or building full-stack starter kits. AI Gateway, AI SDK, Sandbox, and BotID are all components of the Vercel AI Cloud, designed to work independently or as a full stack running at the Edge.

Clone the vibe coding platform, swap models, add your own tools, or connect a database. The repository lives in Vercel's examples folder on GitHub.