Understand Tokens and estimated costs
Interpret member and model consumption, check complete, partial and unknown pricing states, and distinguish estimates from invoices.
Tokens measure model usage. USD amounts are cost estimates based on recorded usage and catalog prices. Neither directly measures work value, and an estimate is not a provider invoice.
Prerequisites and permissions
- Complete the identity and workspace checks in Member usage and visibility. The current Team → Cost analysis entry is for platform administrators; its data remains limited to readable, recorded Runs.
- Usage records and applicable model catalog prices are needed. Missing collection or pricing can make costs incomplete or unknown. Missing prices do not mean free usage.
- Cost review does not require Prompt/Response content. Any necessary content access still requires explicit reveal, authorization and auditing; a failed audit denies the read.
- This guide describes the current Console implementation. It does not establish that your deployment has validated model metering or reconciled provider invoices.
Steps
- In the target workspace, open Team → Cost analysis. Review member, task, Token, cost and share-of-cost columns in Member costs.
- Read “Task” as the number of Runs attributed to that member's Owner identifier. Member cost is the sum of those records' estimated amounts. Its share is relative to the current member aggregate, not a share of a fully reconciled invoice.
- Inspect Cost in the last 7 days separately. The current member table aggregates returned records, while the trend selects the last seven days by date. There is no unified date filter, so the displayed table total need not equal the sum of those seven days.
- For model distribution, open Team → Overview and inspect Model usage distribution, or review model rows under Models & quota. Align record scope first. The model page's “Usage this period” is not automatically the UTC-month total used for quota admission.
- For material differences, ask authorized platform operations staff to inspect the workspace, Run identifier, Owner, model identifier, start time, input/output Tokens, estimated amount, pricing state and missing-price information in the underlying records. The current summary tables do not expose all pricing states; this guide cannot promise a “click the amount to reconcile” workflow.
The cost-page Console route pattern is /w/{workspaceId}/settings?section=team-cost. To change admission rules, read Model enablement and monthly quotas. A cost page does not establish that budget blocking or alerts are enabled.
Pricing states and Token accounting
| Pricing state | Meaning | How to use it |
|---|---|---|
complete | The record's usage is fully priced under the applicable catalog rates | Use for that record's estimated-cost analysis; it is still not a provider invoice |
partial | Only some usage has prices; the known amount is retained | Record missing-price items alongside the amount; do not present the known subtotal as complete cost |
unknown | The record's cost cannot currently be determined | Mark it unknown and check usage and pricing sources; do not treat it as zero dollars |
Input, output, reasoning, cache-read and cache-write Tokens can have different pricing buckets. The platform normalizes inclusion relationships so cache already included in input, or reasoning already included in output, is not counted twice. A single total multiplied by one rate need not equal a Run's estimate.
Prices and release versions can form a Run's pricing snapshot. A later catalog price is not a substitute for the historical snapshot. An unset or zero catalog rate does not mean the model is free. A displayed $0.00 can also reflect unknown or partial pricing, or display precision.
Verification and investigation
- Check a sample: select existing, authorized Run samples, establish that usage is known, and compare estimates using their pricing snapshots and mutually exclusive Token buckets. If pricing is missing, validate only the known portion and retain the
partial/unknownstate. - Check attribution: align workspace, Owner, model and time range before comparing totals. Member tables and all-Run aggregates can also differ because of membership lists or missing ownership. Do not arbitrarily assign unattributed amounts to a person.
- Check invoices separately: follow your finance process to reconcile provider invoices, discounts, contracts, taxes and other charges. A Console estimate does not establish invoice reconciliation.
- Handle gaps: if Tokens exist without cost, check price coverage and usage state. For amount differences, check buckets, snapshots, model identifiers, time and record completeness. Report failed loading or collection as a gap, not as savings.
Limits
Costs cover only the readable, recorded Runs and known priced portions in scope. Private development_machine Runs remain owner-only. Summary tables and CSV exports do not fully express estimation coverage and cannot be treated as the enterprise's complete bill.
Monthly cost quotas use recorded estimates for Task-creation admission. They are not real-time debits or provider-enforced hard budgets. There is insufficient evidence for a unified cost-threshold alert tutorial; this guide promises neither automatic notifications, real-time circuit breaking nor a fixed savings percentage.