Cut your AI bill in half - Batch API

I caught it in one of my own scheduled jobs. Every night at 2am, a task refreshes a competitive intelligence dashboard: summarize the scraped pages, write the notes nobody reads until morning. Every token billed at real-time prices, and not a single human waiting on the response.

That is the question most stacks never ask: is anyone waiting on this response right now? Synchronous APIs become the default everywhere because the first feature you build is interactive, and every job after that inherits the habit.

The big providers now run two lanes for the same models. Real time, where you pay a premium for immediacy. Batch, where the same model returns the same quality later and costs half: Anthropic and OpenAI both discount batch a flat 50 percent, and most jobs come back well inside 24 hours, often within the hour. It is courier versus ground shipping. You are paying for speed, not intelligence.

The jobs that belong in the cheap lane are obvious once you look: evals against thousands of prompts, classifying a backlog of tickets, embedding a document corpus overnight, the nightly digests and morning dashboards that are deliberately scheduled off-peak anyway.

The sorting rule fits in one sentence. A person is actively waiting on this specific response: real time, pay for it. Nobody is waiting: batch it, same output, half price.

Two cautions before you re-route everything. Batch has a 24-hour ceiling, not a guarantee, so nothing on a critical path should depend on it without a fallback. And chase the savings where the spend actually is: halving a serious evals bill matters, doing surgery on pennies does not.