Usage and cost analytics
Last updated September 15, 2026
What the usage and cost pages each answer, the available breakdowns, and how they relate to billing.
The console provides two analytics pages, answering different questions:
| Page | Answers |
|---|---|
| Usage | How much was called — request counts and the distribution across token types |
| Cost | How much was spent — how the amount splits across models, keys, and billing components |
Both read from the same daily aggregates, so long time ranges stay fast.
Usage
Metrics
| Metric | Meaning |
|---|---|
| Total tokens | The sum of input, output, cache read, and cache write |
| Input / output tokens | Prompt tokens and tokens the model generated |
| Cache read / write tokens | Tokens served from and written to the cache |
| Total API requests | Number of calls in the range |
Upstream providers have no uniform "total" field in their usage reports, and their definitions differ, so the total is computed by the platform as the sum of the four figures above.
Breakdowns
- By model — which models consumed how many tokens
- By API key — how much each key consumed, the primary way to attribute usage to a project
- By workspace — visible in the organization view, for comparing departments
- By member — filterable by member in the workspace view
Cost
Metrics
| Metric | Meaning |
|---|---|
| Total cost | All spend in the range |
| Input / output cost | The amount for each of the two main components |
| Other cost | Cache reads and writes, per-request charges, and the remaining components |
| Average per request | Total cost ÷ request count |
| Average per million tokens | Total cost ÷ total tokens, useful for comparing the real cost of different models |
Breakdowns
- By model — which models the money went to
- By billing component — the share taken by input, output, cache read, cache write, reasoning, per-request, and the rest
- By API key — attribution to a project
- By workspace — departmental split in the organization view
"By billing component" is usually the first place to look when optimizing cost: a high share of cache writes often means a low cache hit rate, while a high share of reasoning tokens suggests the reasoning effort is set higher than the task needs.
Relationship to billing
The analytics pages and the billing page use the same definitions: amounts are actual charges.
Anything not charged does not appear here, including calls that returned an upstream error and attempts that failed during routing failover — those costs are absorbed by the platform. See billing overview.
Visibility
| Role | Scope |
|---|---|
| Organization owner / admin | Organization-wide, splittable by workspace and member |
| Organization finance | Organization-wide amounts |
| Workspace admin | All of their workspace |
| Workspace member | Their own only |
Attribution is only as fine as your keys
Usage and cost are reported per key, and requests made with the same key cannot be separated afterwards. To distinguish purposes, create a separate key for each from the start.
Calls made with a shared key also record the actual operator, so the playground and similar contexts can still be attributed to a person; calls made through the API, however, can only be attributed to a key.
Related
- Call logs — the full record of a single call
- Billing overview — how charges arise
- API keys — splitting keys by purpose