How much is cached input?
Meta lists Wonder input at $4 per million tokens and cached input at $0.20. The cached rate is 95% lower, or one twentieth of the normal rate. This discount applies only to the cached input portion. Output and search rates stay the same. Official rates and caching notes ↗.
The pricing notes describe a matching cached prefix and a response field, cached_tokens, that identifies the input served from cache. Similar meaning, repeated subject matter or your own local cache does not by itself prove an API cache hit.
One million total input tokens
| Cached share | Uncached charge | Cached charge | Combined input cost |
|---|---|---|---|
| 0% | $4.00 | $0.00 | $4.00 |
| 25% | $3.00 | $0.05 | $3.05 |
| 50% | $2.00 | $0.10 | $2.10 |
| 75% | $1.00 | $0.15 | $1.15 |
| 100% | $0.00 | $0.20 | $0.20 |
These are mathematical scenarios, not promised hit rates. At 50% caching, 500,000 tokens are charged at each rate. Total input is $2.10, not $4 plus an extra cache fee.
Calculate savings accurately
Input savings = cached tokens × ($4 − $0.20) / 1,000,000
= cached tokens × $3.80 / 1,000,000If 80,000 of 100,000 input tokens are hits, input costs $0.08 uncached + $0.016 cached = $0.096. Without hits it would cost $0.40. The input saving is $0.304 per request.
Add 10,000 output tokens and the request totals $0.296 instead of $0.60 before other tools. The total-bill discount is smaller than the input discount because output remains full price.
A cache slider is a planning assumption. It does not enable caching, send an API request or guarantee a repeated prefix will receive a discount.
A simple measurement checklist
- Record total input and the reported cached portion across representative requests.
- Keep billable output and tool-query counts separate.
- Estimate the cache share for the same workload and period you plan to budget.
- Recheck after changing prompt structure, documents or request patterns.
We do not assert a guaranteed cache lifetime, minimum prefix or behavior across providers. Those details can differ. Consult Meta’s prompt-caching guide ↗ and your provider’s implementation notes.
Use caching in a budget
In the calculator, enter total input, then set the cached percentage supported by your data. Leave it at zero for a conservative case if you have not measured hits.
Compare the complete API bill and Spark Standard pricing using the same assumptions. Caching is one cost factor alongside output length, tools, retries and successful results.