2026-10-28-caching-benchmark

{
  "id": "K01",
  "title": "Cost Optimization on the Claude API",
  "publisher": "Anthropic / Ben Lehrburger",
  "url": "https://platform.claude.com/cookbook/cost-optimization-cost-optimization",
  "publishedOn": "2026-08-09",
  "retrievedOn": "2026-10-28",
  "locator": "Baseline run; Model selection; Putting it all together, carried-over configuration table",
  "scope": "Saved API runs on 10 synthetic insurance-claim evaluation cases; Opus 5, high effort. Token-accounted API costs include input, output, cache writes and reads.",
  "limitation": "One baseline trial and two cached trials on the same 10 unique test cases, not a customer production bill. Request counts differ. 10/10 correct in each run does not establish broader quality equivalence. Excludes human review, infrastructure and operations. Historical Opus 5 rates, not the separate Opus 5.5 price example.",
  "results": [
    {
      "configuration": "No cache",
      "trial": 1,
      "usd_per_claim": 0.2906,
      "correct": 10,
      "claims": 10,
      "requests": 36
    },
    {
      "configuration": "Cached",
      "trial": 1,
      "usd_per_claim": 0.1498,
      "correct": 10,
      "claims": 10,
      "requests": 34
    },
    {
      "configuration": "Cached",
      "trial": 2,
      "usd_per_claim": 0.1512,
      "correct": 10,
      "claims": 10,
      "requests": 35
    }
  ],
  "calculation": {
    "cached_mean_usd_per_claim": 0.1505,
    "reduction_percent": 48.21059876118377,
    "method": "Unweighted mean of the two published, rounded cached trial costs; reduction against the published baseline. Display as about 48%. No volume extrapolation."
  }
}