I Routinely Send the Boring 80% to the Cheap Model. You Should Too.

Claude Haiku 5.5 at roughly 75% cheaper than its predecessor changed how I spend money on AI, and not in the way you'd think.

Most of my calls are high-volume and low-stakes: tagging, summarizing, routing, quick sub-agent steps. Those don't need the biggest model, and paying top dollar for them is how a bill balloons.

So I route. The cheap, fast model handles the 80% that's mechanical, and I bring out the heavy model only for the judgment calls where quality actually shows. The result reads the same to the end user and costs a fraction.

Anthropic's adjustable thinking effort and OpenAI's tiny routing models point the same way: the future isn't one giant model, it's a fleet with a dispatcher.

Stop paying flagship prices for the work a fast small model does fine. The dispatcher is the real upgrade.