Quick answer
Local session data is written by the tool as you work and gives per-turn detail, but it can miss other machines and nothing enforces it. Provider-reported usage is aggregated by the vendor and is the authority for limits and billing. Use local data to understand, provider data to decide.
Side by side
| Local | Provider | |
|---|---|---|
| Granularity | Per turn or session | Per account or seat |
| Latency | Immediate | Minutes to hours |
| Coverage | This machine only | Everywhere you signed in |
| Authority | Your record | The bill and the limit |
| Privacy | Stays on disk | Held by the vendor |
Why they disagree
- The provider includes usage from web, desktop, and other machines.
- Local logs may exclude cached inputs or reasoning tokens.
- Plan allowances are priced differently from raw tokens.
- A reset can change the provider number without new usage.
None of these are bugs. They are different measurements of the same activity.
How to use both
Track local data to see which task, model, or project consumed the tokens. Read the provider panel to decide whether you can continue. UsageAtlas shows both in one window for the providers that expose them, and why usage numbers differ explains the arithmetic.
Questions, answered
Which usage number is correct?
Both. Local logs are detailed but partial; provider figures are authoritative for limits and billing. Compare them for different purposes.
Why is provider usage higher than local?
It includes other machines, web and desktop sessions, and usage the local tool did not record.
Can local data replace a dashboard?
For understanding cost and projects, often yes. For knowing whether you are about to be blocked, use the provider's reported number.