PURE · AI · Cost & efficiency
Where the AI spend actually goes — by model, surface and cron — with a monthly budget, projection, and the efficiency moves that cut it most. See it, cap it, route cheap work to cheap models.
Sample spend — wire mcp_ai_spend for liveIf the default model is the expensive one, most spend is here. Route the cheap 80% (classification, extraction, drafting, summaries) to a fast model and reserve the reasoning model for hard tasks.
| Model | Calls | Tokens | Spend (MTD) | Share |
|---|
Surfaces and crons ranked by spend. Scheduled LLM loops are the silent money sink — an hourly job on the expensive model adds up fast.
| Driver | Type | Calls/day | Spend (MTD) |
|---|
Ranked by impact. These are the levers that stop the overspend — most are backend flips, no rebuilds.
Turn each cost lever on/off. Defaults are the money-savers. Changes persist to pure_ai_controls; pure-chat-api + the worker read it every run, so a toggle takes effect on the next call.