# Prompt caching: absolute API cost evidence

28 October 2026 · replacement for the main Care Access percentage-only example

Anthropic’s [9 August 2026 cookbook](https://platform.claude.com/cookbook/cost-optimization-cost-optimization) reports saved runs on ten synthetic insurance claims using Opus 5 at high effort. Its final configuration table carries forward the earlier measured runs; these are not additional replications.

| Configuration | USD per claim | Correct | API requests |
|---|---:|---:|---:|
| No cache, one trial | 0.2906 | 10/10 | 36 |
| Cached, trial 1 | 0.1498 | 10/10 | 34 |
| Cached, trial 2 | 0.1512 | 10/10 | 35 |

The cached mean is $0.1505; the reduction against $0.2906 is 48.21%, displayed as about 48%. Bars share a zero origin and use those exact costs. The small evaluation does not demonstrate production accuracy or isolate every stochastic difference. Costs cover the model API, including cache writes/reads and output, not the complete application. No latency result is inferred. No annual or 10,000-task bill is claimed.

Slide 14 shows the absolute comparison and the synthetic-test label. Narration explains the range, sample and scope. Slide 15 then explains stable-prefix reuse with a separately labelled current Opus 5.5 price example. Care Access remains historical source material; its absolute baselines remain undisclosed.

Source register: `../03-sources/2026-10-28-caching-benchmark.json`. Chart inputs: `../02-data/2026-10-28-caching-benchmark.csv`.