Model Routing vs Single Model

Route three pull requests

Here are three pull requests with bugs planted in them. The smallest changes a price formatter and the riskiest changes how refresh tokens work. Pick one, then switch between sending everything to Opus and letting the router decide, and see what each review caught.

Treat this as a worked example. The diffs, planted bugs, router rules and prices are real, but the reviews, timings and output token counts are made up to show how the comparison works. Every review used the same prompt with effort set to high, and you can run the same comparison on your own pull requests with the code below.

What routing saves at your volume

Three pull requests don't tell you much about your bill, so let's scale it up. Pick how many pull requests your team opens a month and what most of them look like.

Each pull request is priced like the example of its size above, with the path rule on. The prices are Anthropic's published API rates; the token counts come from the made-up runs, so treat the totals as a rough guide and plug in your own.

How much code routing adds

Here's everything behind both approaches, using the Anthropic TypeScript SDK. The routed version is the same call with a router in front of it, and the highlighted lines are what it adds.

import Anthropic from '@anthropic-ai/sdk' const client = new Anthropic() const SYSTEM = `You review pull requests. Report every problem that should be fixedbefore it merges: bugs, security issues, performance problems and missing edgecases. Leave out style preferences.` const MODEL_FOR = {  simple: 'claude-haiku-5-5',  medium: 'claude-sonnet-5-5',  complex: 'claude-opus-5-5'} // Changes here always get the best model, however small they areconst RISKY_PATH = /(^|\/)(auth|payments|billing|migrations|security)\// function classify(diff: string): keyof typeof MODEL_FOR {  const files = [...diff.matchAll(/^\+\+\+ b\/(.+)$/gm)].map(match => match[1]!)  const changedLines = diff.split('\n')    .filter(line => /^[+-]/.test(line) && !/^(\+\+\+|---) /.test(line))    .length   if (files.some(file => RISKY_PATH.test(file)) || changedLines > 300) {    return 'complex'  }  if (changedLines > 40 || files.length > 2) {    return 'medium'  }  return 'simple'} export async function reviewPullRequest(diff: string) {  return client.messages.create({    model: MODEL_FOR[classify(diff)],    max_tokens: 16000,    system: SYSTEM,    output_config: { effort: 'high' },    messages: [{ role: 'user', content: `<diff>\n${diff}</diff>` }]  })}

The router only reads the diff, so it doesn't add an API call or any waiting. What can go wrong is a dangerous change that looks small, like the refresh token pull request above when the path rule is off.