Vercel adds Claude Opus 5 to AI Gateway with fallbacks

Key Takeaways

• Claude Opus 5 handles multi-file refactors and end-to-end features without leaving stubs
• Model fallbacks let teams route around safety classifier blocks automatically
• Reasoning effort is configurable from minimal to xhigh, with 'medium' offering the best quality-to-latency ratio

Claude Opus 5 is now live on Vercel's AI Gateway, giving developers a single API to run Anthropic's newest model alongside automatic fallbacks, spend tracking, and access to competing providers. The update targets teams building coding agents and multi-agent workflows that need long-horizon task completion.

Opus 5 improves on earlier Opus versions in a specific way: it finishes what it starts. Where previous models might leave stubs or placeholders mid-refactor, Opus 5 completes full tasks across multiple files. For engineering teams running coding agents, that difference cuts review cycles and manual cleanup.

What does Opus 5 actually do better?

Anthropic positions Opus 5 for agentic coding, the kind where a model handles larger refactors, multi-file features, and end-to-end feature work without human hand-holding between steps. The model performs well at low and medium reasoning effort, meaning teams can get quality outputs without burning tokens or waiting on high-latency responses.

Vision capabilities got a bump too. Opus 5 handles charts, documents, diagrams, and UI replication better than its predecessors. It can analyze, crop, and verify visual work using tool calls, which matters for teams automating design-to-code pipelines or document extraction.

The model also coordinates teams of subagents effectively. If you're building multi-agent workflows where one model dispatches tasks to others, Opus 5 is built for that orchestration role.

How do reasoning levels work in AI SDK 7?

Reasoning is on by default. In AI SDK 7, you set the effort level with a top-level reasoning option. The settings range from minimal to xhigh, or none if you want to disable it entirely.

import { streamText } from 'ai';

const result = streamText({
  model: 'anthropic/claude-opus-5',
  prompt: 'Implement the multi-file refactor described in this issue.',
  reasoning: 'medium',
});

The medium setting hits a sweet spot: good output quality at a fraction of the tokens and latency you'd burn on higher settings. For most coding agent use cases, that's the default worth starting with.

Why do model fallbacks matter for Opus 5?

Like Fable 5, Opus 5 ships with elevated cybersecurity safeguards. That's good for preventing misuse. The downside: legitimate security work, like auditing scripts for vulnerabilities, can occasionally trigger the safety classifier and get blocked.

Vercel's solution is model fallbacks. You list backup models in a models array under providerOptions.gateway, and the gateway tries them in order when the primary model fails. If Opus 5 blocks your security audit request, the gateway routes it to Opus 4.8 or Sonnet 5 automatically.

import { streamText } from 'ai';

const result = streamText({
  model: 'anthropic/claude-opus-5',
  prompt: 'Audit this script for vulnerabilities.',
  providerOptions: {
    gateway: {
      models: ['anthropic/claude-opus-4.8', 'anthropic/claude-sonnet-5'],
    },
  },
});

This fallback pattern works across every AI Gateway API format. It's a practical fix for the false-positive problem that strict safety classifiers create.

Fast mode and provider options

Opus 5 supports a fast mode for latency-sensitive applications. Enable it by passing speed: 'fast' in the anthropic provider options. The tradeoff: faster responses, potentially lower quality on complex tasks.

Multiple providers serve Opus 5 through AI Gateway. To see every provider, along with per-provider pricing, uptime, throughput, and latency, you can call the model endpoints API directly.

Opus 5 is also compatible with Zero Data Retention for teams with compliance requirements around data handling.

Setting up Opus 5 in your coding agent

Vercel provides a CLI command to wire Opus 5 into existing coding agents. Running vercel ai-gateway coding-agents setup detects agents on your machine, provisions an API key, and writes the config. The setup guide covers integration with Claude Code specifically.

The unified spend tracking alone justifies routing through AI Gateway for teams running multiple models. You get one bill, one dashboard, and consistent fallback behavior across providers.

Frequently Asked Questions

Is Claude Opus 5 available through Vercel AI Gateway now?

Yes, Opus 5 is live on AI Gateway as of July 2026, accessible through the AI SDK 7 with the model identifier 'anthropic/claude-opus-5'.

What reasoning effort should I use for coding tasks?

Start with 'medium'. It produces quality outputs at a fraction of the tokens and latency of higher settings. Use 'xhigh' only for complex tasks where accuracy matters more than speed.

How do I prevent Opus 5 from blocking legitimate security work?

Use model fallbacks. List backup models like Opus 4.8 or Sonnet 5 in providerOptions.gateway.models, and the gateway routes blocked requests to them automatically.

Does Opus 5 support Zero Data Retention?

Yes, Opus 5 is compatible with Zero Data Retention for teams with compliance requirements.

Can I see pricing for different Opus 5 providers?

Yes. Call the model endpoints API at ai-gateway.vercel.sh/v1/models/anthropic/claude-opus-5/endpoints to see per-provider pricing, uptime, throughput, and latency.