Usage and cost
Separate the usage in one model response from the complete cost of answering the request.
Response usage
Read the usage fields returned by your model API when present. Input, output and cached-token fields vary by provider. Missing usage is unknown, not zero.
Total cost
For an evaluation, include all model work, tlc processing, checks, retries and fallbacks. A response’s token count alone does not establish the complete cost or a saving.
Call records
Where your setup exposes call records or receipts, use them to trace usage and diagnose requests. Collecting these records is handled by the system; your application’s normal request does not require a separate reporting call.
Compare like for like
Record the model and pricing date, distinguish provider charges from estimates, and evaluate quality alongside cost. See Evaluation.
The Transductor is in early access. Join the waitlist.