NexoToken technical guide center
Codex tokens, prompt caching, and cost verification
Reconcile input, output, and cache-hit tokens instead of estimating cost from a total or old price.
A Codex cost cannot be verified by multiplying total tokens by one price. Fix the request time and exact model, separate input, output, and cache-hit tokens, then apply the current price for each component. Check automatic retries and client-added context before explaining any balance movement.
For users reconciling a Codex task, checking prompt-cache hits, or explaining differences between token counts and recorded cost.
Last verified
Steps
Record request time, model, and a safe record identifier.
Read input, output, and cache-hit tokens separately.
Check current units and update time for the exact model.
Calculate each component and compare recorded cost.
Inspect retries, hidden context, and price timestamp for material differences.
Safe example
Verified cost = input component + output component + cache-hit component
Use the current price for each part; do not treat points as fiat currency.Common errors and troubleshooting order
- Combining all tokens hides distinct pricing.
- A related model version can have a different price.
- Automatic retries create additional real requests.
- A fixed price in an old article is not current evidence.
Completion checklist
- Model and time match one record.
- Three token categories were checked separately.
- Current public prices were used.
- Evidence excludes prompts, credentials, and personal data.
Limitations and boundaries
- Cache hits depend on request structure and model capability, not just a client toggle.
- Examples explain the method; actual usage and current prices remain authoritative.