A 90% cached-input discount sounds like a 90% saving. It is not—not on the first reuse, and not after the cache write is counted.
OpenAI's current GPT-5.6 model pages list cached input at one tenth of ordinary input, but cache writes at 1.25 times the ordinary rate. That makes the first request more expensive. The useful question is how many times the same prefix must be used before the write pays for itself.
It pays back on use two
Let one ordinary pass over a reusable prefix cost one unit. Without caching, N uses cost N units. With caching, the write costs 1.25 units and each later hit costs 0.10:
uncached = N
cached = 1.25 + 0.10(N - 1)
saving = 90% - 1.15/N
The first use costs 25% more. At two uses, the total falls from 2 units to 1.35: a 32.5% saving. Five uses save 67%; ten save 78.5%. The realized saving approaches 90% as reuse grows, but a finite batch never quite reaches the headline percentage.
The model tier changes dollars, not the break-even
For a one-million-token prefix used twice, the documented rates produce the same 32.5% percentage saving on all three GPT-5.6 API tiers:
| Model | Two uncached uses | Write plus one hit | Saving |
|---|---|---|---|
| Sol | $10.00 | $6.75 | $3.25 |
| Terra | $4.00 | $2.70 | $1.30 |
| Luna | $0.40 | $0.27 | $0.13 |
Those figures come from the current official pages for Sol, Terra, and Luna. Output is excluded because caching changes the input side of this comparison.
The calculation is the easy part
This is a documented cost model, not a production cache test. It assumes the prefix is eligible and each later request hits. I did not pay for API calls, so I did not measure eviction, hit rate, latency, or how often a real prompt changes enough to miss.
That boundary matters more than another decimal place. A stable system prompt reused across a batch can cross the break-even on its second use. A prompt rewritten every time may never create the hit the spreadsheet assumes.
The honest planning number is therefore not “90% cheaper.” It is the reuse curve multiplied by the share of prefix tokens that actually hit. The headline discount is the ceiling; the workload decides how close you get.
