Post

The old model was cheaper batch processing. You could simply move your LLM calls to an async model, wait 12-24 hours, and save 50%. This allowed the providers to load shift. That doesn't work in the agentic world though, where a session can have dozens or hundreds of calls that have to go in order.