Prompt Architect

Questions

Prompt cost, in plain terms

• How do you work out what a prompt will cost before you send it?

  Cost is token count times the model's input price, plus the reply at its output price — but the money hides in the details. This tool counts the blocks you wrote, the chat-template scaffolding the API wraps around them, and the tokens your tool definitions occupy, then prices all of it against the current rate for every model in the picker. Tool definitions in particular are billable input that people routinely forget they are paying for on every single call.

• What is prompt caching, and when does it actually save money?

  A provider can hold the unchanging front of your prompt — a long system prompt, a fixed document — and charge much less to read it again. The catch is a floor: at Anthropic a block has to reach 4,096 tokens before it can be cached at all, at Google 1,024. Writing the cache also costs more than a plain read (1.25× at Anthropic) before it starts paying back at 0.10×. A block under the floor earns nothing, so this tool shows zero savings for it rather than a discount that will never arrive.

• Does the order of the blocks in a prompt matter?

  For caching it is the difference between paying once and paying on every call: whatever you want cached has to sit at the front and stay byte-identical between requests. Put a timestamp or a user name above your long system prompt and the cache misses every single time. Roles matter for a separate reason — system for standing rules, context for retrieved material, user for the actual request — because models weight them differently, and burying an instruction inside pasted context is the usual way it gets ignored.

• What is the Batch API, and when is it worth using?

  The same model at roughly half price, in exchange for giving up the immediate answer: you submit a batch and collect the results later. Anthropic and OpenAI both document a 50% discount, so this tool applies it for them; Google documents none, so nothing is applied there rather than a guessed figure. It is worth it for anything with no human waiting on the other end — evaluation runs, bulk classification, offline generation.