Claude Sonnet 4.5 lands on Vercel's AI stack

Vercel has added support for Anthropic's Claude Sonnet 4.5 across its AI infrastructure. The model is now available through the AI Gateway and fully supported in the AI SDK. Vercel has been testing it internally in v0, in its Next.js build pipelines, and in a new open-source Coding Agent Platform template.

Early results point to meaningful gains in both design sensibility and code quality, particularly when building and linting Next.js applications. The model builds on Anthropic's reasoning and coding strengths, and Vercel sees it as a fit for developer workflows where AI agents plan, execute, and safely ship changes inside repositories.

Measuring the gains in agentic coding

Vercel's internal evaluations show Claude Sonnet 4.5 outperforming its predecessor on core agentic coding tasks. In Next.js builds, the model achieves higher first-pass success rates with fewer missing dependency errors; it correctly infers which packages are needed and adds them to package.json before attempting a build. Linting results show stronger adherence to Next.js and ESLint defaults. When implementing features, the model handles framework-native components more reliably, such as correctly importing <Image> from next/image.

Inside v0, Vercel observed several improvements:

  • Cleaner component structures, better Tailwind usage, and more accessible UI patterns in frontend design
  • Better handling of nuanced prompts, with the model finding relevant context faster and avoiding derailment by unclear requirements
  • For complex applications like 3D worlds, Claude Sonnet 4.5 often succeeds in a single attempt, coordinating moving parts and interactions without extended back-and-forth
  • When a prompt is ambiguous, the model asks for clarification rather than spending tokens on an irrelevant guess

Taken together, these capabilities support coding agents that execute full task chains: parse a request, generate code, verify with builds and tests, and ship the result.

A reference platform for running agents

To demonstrate these capabilities, Vercel has released a Coding Agent Platform template that defaults to Anthropic's Claude Code agent running Claude Sonnet 4.5. The template supports a multi-agent system that can run Claude Code, OpenAI Codex CLI, Cursor CLI, or opencode.

How it works: after connecting a GitHub organization, you select a repository and define a task, such as adding dark mode, linting a project, or creating a new API route. The agent plans, executes, and commits changes automatically. Each task runs in an isolated sandbox: the platform installs dependencies, runs CLI commands with the selected model, and pushes changes to GitHub for review. The UI, built with Next.js and Tailwind CSS, streams commands, builds, and commits in real time, with state persisted in Postgres for task history and management.

Built on the Vercel AI Cloud execution layer

The Coding Agent Platform template relies on the Vercel AI Cloud, with each layer making background agents viable in production. At the base, Vercel Sandbox provisions secure, short-lived environments for installs, builds, and tests, keeping tasks reproducible and isolated from the rest of a project. Full command-line access lets agents clone repositories and use complete agentic coding toolchains.

Sandboxes scale on demand, spinning up in parallel as tasks are created. Built on Fluid compute with Active CPU pricing, agents can idle without incurring charges, with costs accruing only during active compute. AI calls run through the AI SDK and AI Gateway; for instance, when an agent starts a task, the system generates a Git branch name based on the task, repository, and context.

The result is an integrated execution layer connecting model reasoning with scalable, safe execution, rather than assembling separate tooling.

Open-sourcing and getting started

The Coding Agent Platform template is open source, published under the same vision as Vercel's Open SDK strategy. You can fork and audit the code, plug in new agents or domain-specific tools, customize commit logic, or use it as a reference implementation while taking advantage of AI Gateway, AI SDK, and the Vercel AI Cloud. Documentation, examples, and a contributor guide will accompany the release.

Claude Sonnet 4.5 is available today through AI Gateway and AI SDK for immediate testing in the playground, and the Coding Agent Template is ready to deploy against live repositories.