Quick answer
To compare Cached tokens results, keep input, format and settings fixed and change one variable. Prompt caching can cut input cost, but hit rate depends on repeated context and cache policy.
Why it matters
Changing several variables at once makes results unattributable.
Only hits are cheaper.
Step by step
- Measure input, output and retry tokens per user/task.
- Separate cache, batch, embeddings and fine-tuning costs.
- Use 30 days of real requests for cost per user, then margin.
- Model prices change; baseline from official pricing pages and record the date.
Common mistakes
- Estimating from input tokens only; output often costs more.
- Forgetting retries, failures and tool calls.
- Assuming cache hits by default.
- Using stale prices for long-term margin.
Checklist
- Input/output/cache/retry tokens are tracked separately.
- Batch/async fit and latency requirements are clear.
- Embeddings/fine-tuning/image/audio are separate.
- Cost per user and margin use real usage.
- Price date and fallback model are documented.
Key facts
- Only hits are cheaper.
- Caches have TTL.
- Not all models support.
FAQ
How do I compare two Cached tokens results?
To compare Cached tokens results, keep input, format and settings fixed and change one variable. Prompt caching can cut input cost, but hit rate depends on repeated context and cache policy.
How long does it take?
Usually 1-3 minutes; for intermittent issues, run 30 seconds and retest elsewhere.
Is data uploaded?
Runs locally and does not upload video/audio; extensions may still read the page.
Source: OpenAI