Cache Writes Cost More, Not Less, on GPT-5.6

The mistake

Treating prompt caching as a pure discount on GPT-5.6-family models, the way it behaved on every model before them. On the older generation, writing a new prefix into the cache never cost anything extra — you paid the ordinary input rate to send it, and any later hit against that prefix was simply cheaper. Carry that assumption into GPT-5.6-sol, Terra or Luna and you'll under-budget a workload that creates a lot of fresh cache prefixes without reading many of them back.

Why this happens

Starting with the GPT-5.6 family, a cache write is its own priced line item, billed above the plain uncached input rate rather than at it. The pricing tables for every earlier model simply don't have that column — cache writes were free by omission, not by a stated zero rate — so there was nothing to notice changing. GPT-5.6 introduces the column, and it's a markup, not a discount, which is the part that catches people who migrated a caching-heavy workload up a model generation without re-reading the pricing page for the new model specifically.

Why it matters

Caching only pays for itself once a prefix gets reused enough times to make up for the extra cost of writing it in the first place. On a model where writes were free, almost any reuse was a win. On GPT-5.6, a workload that writes a lot of prefixes it barely re-reads — think short-lived sessions, or a prompt_cache_key that gets rotated too aggressively — can end up paying more overall than it would with caching disabled entirely, and the bill won't say why; it'll just look larger than the caching feature was supposed to make it.

The fix

Before assuming a caching strategy that worked on an older model still helps on a GPT-5.6-family one, work out the real break-even for that model: how many later hits a given prefix needs before the write pays for itself. The site's prompt-caching reference computes this directly from the published write and read multipliers rather than leaving it as a guess, and it's worth checking per model, since the multipliers themselves are model-specific.

See also

The Prompt-Caching Savings Calculator models this exact break-even for the model and workload you're actually running, instead of the generic version above.