Member-only story
I turned prompt caching off on a Claude Opus 5 assistant last week and the input bill dropped 20% — from $4.61 to $3.69 for the same 41 requests over the same 8 hours.
That is not how caching is supposed to work. Every provider doc leads with the same number: cached reads cost 90% less than fresh reads. A 10x discount. Free money for anyone with a stable system prompt.
The catch is that since July 9, 2026, when GPT-5.6 went GA across Sol, Terra and Luna, cache writes are billed on both of the two biggest APIs. OpenAI replaced its old free implicit writes with explicit breakpoints, a 1.25x write premium and a 30-minute TTL, which lands it almost exactly where Anthropic has been all along. Prompt caching stopped being a one-way bet and became a wager: you pay a premium up front for a discount you might never collect.
If your cache entry expires before anyone reads it, you paid the premium for nothing. My assistant had 12 minutes between turns and Claude’s default TTL is 5 minutes. Every single request was a cold write at 1.25x. Caching was a 25% surcharge dressed up as a 90% discount. I wrote a 200-line zero-dependency script to price out all four providers properly. Three findings I did not expect: