Tokens need categories
A total token count is useful for volume, but insufficient for cost. Providers can price input, cached input, and output differently. The execution layer must retain the categories supplied by the provider and apply rates only when the required details are available.
Cost belongs beside the run's model, version, attempts, and latency. A short run with expensive output behavior and a long tool-heavy run can have different usage patterns. Run-level context prevents a single aggregate number from hiding those differences.
Snapshot the rates
Dynamic pricing lookup makes historical results unstable. Runovio uses reviewed, code-owned pricing snapshots with effective dates. A run keeps the estimate produced by the snapshot in effect when it executed, even if a later release adds new rates.
The calculation uses integer micros rather than binary floating point. That keeps rounding explicit and supports boundary tests. Updating the catalog requires checking the official provider pricing source, recording a new date, reviewing every input, cached-input, and output rate, and preserving older snapshots.
Unknown does not mean free
When the selected model is absent from the catalog, or a required usage category has no trustworthy rate, the correct value is unavailable. Reporting zero would imply that the work was free and can distort operational decisions. Runovio therefore stores nullable estimated cost and displays missing estimates as unavailable.
An estimate also does not replace provider billing. OpenAI bills the account behind the workspace's own API key. Account-specific terms, discounts, taxes, and adjustments may differ from a catalog estimate. Runovio's number is an observability signal for comparing runs under a known snapshot.
Use cost with tokens, attempts, and traces to investigate changes. A prompt revision may increase output length; a tool loop may add model rounds; a retry may repeat paid work. Those causes become visible only when cost is connected to execution. Read the cost tracking feature, the cost monitoring use case, and the traces guide.