Cost notes from building beforeyouship

Cost per MAU is the easiest number to compute and the least useful. The one that decides your pricing is the request count at which a single seat stops being profitable — and on a flagship model it's lower than you'd guess.

Batch is exactly half price on every provider that offers it — and it's the smallest of your three cost levers. It doesn't stack with caching the way it looks, and a tier drop beats it outright.

Memory looks like a 2× tax — it's actually quadratic. Every turn re-sends the transcript, so long conversations carry the bill. Caching collapses it back to ~1.3×.

Route easy requests to a cheap model and escalate the hard 20% — roughly half the bill, same quality floor. The catch: one number decides whether it pays off.

Your cost ranking is a snapshot at one retry rate. Because retries bill uncached tokens, the cheaper model can flip — the new sensitivity tripwire shows exactly when.

A single agent task fans out to 5–12 model calls — landing ~13× a chatbot. Step count and cascading retries are the two cost drivers nothing else has.

RAG looks like one feature but it's 3–5 LLM calls per question — and with bigger prompts the bill lands ~6× a simple chatbot. Where the money goes, and how caching claws it back.

The discount is ~90% — but a write premium and your real hit rate decide whether caching cuts 60% off your input bill or nothing.

beforeyouship now runs a remote MCP server. Three tools put the full cost model — retries, caching, cascading agent failures — one tool call away from your editor.

Most teams underestimate by 2–3×. Here's the real math — retries, caching, and growth included — modeled across six models.