What to Verify Before Buying Tokens from a Supplier

Start with pricing: ask for input and output prices per million tokens, separately, per model. A single blended number hides where your workload actually spends.

Then metering: confirm that every response returns its own token counts and billed amount, and that failed or rate-limited requests are not charged. Without this, an invoice cannot be reconciled.

Then the terms: request written availability and first-token latency targets, the measurement method, incident response times and the billing dispute window. Targets you can compare beat a marketing tier label.

Finally, the data terms: where inference runs, how long logs are retained, and whether your prompts are used to train shared models. These are contractual questions, and they should be answered before the first production key is issued.