The four tabs
Cost figures are shown at full precision rather than rounded to cents, at fractions of a dollar per million tokens, the digits past the cent are where your actual spend lives.
Why there are no per-request logs
Analytics are deliberately day × model aggregates: the platform does not keep a browsable log of individual requests, and nothing about your prompts feeds these numbers. Whether the network retains prompts and completions at all is your organization’s choice: switch Store prompts off on the Settings page and they are not persisted, full privacy (see the terms of service for the data-handling details). For request-level debugging, work from your side of the wire:- Log the
request_idfrom error responses and theidfrom successful ones, support can trace a specific request from those (see error handling). - Record the
usageobject your application receives per response, it is the same accounting that rolls up into these aggregates.
What’s worth watching
- Cached-input share (Caching tab): if it’s low for an agentic or chat workload, restructure prompts per the prompt caching guide, it’s usually the cheapest optimization available.
- Cost by model (Cost tab): a surprising leader often means a workload is running on a bigger model than it needs, see Models.
- Output-token share (Usage tab): reasoning models bill their thinking as output tokens, so a jump here after adopting them is expected, tune with
reasoning.effort.
Every member of your organization can view Analytics, it has no billing controls, so it is safe to share broadly as the team’s common picture of usage and spend.