> For GPT‑5.6 and later models, cache writes are billed at 1.25x the model’s uncached input rate, while cache reads continue to receive the 90% cached-input dis

> For GPT‑5.6 and later models, cache writes are billed at 1.25x the model’s uncached input rate, while cache reads continue to receive the 90% cached-input discount. Not them joining Anthropic with this bullshit. * Caching infrastructure is already a leaky abstraction over a feature that is not as reliable or debuggable to the end user as it should be, charging for the 'privilege' of interacting with it is really annoying. (* for reference on 'this bullshit': ChatGPT previously didn't require anything special for a basic level of caching. Unless you wanted extended cache times, it'd just "do the right thing" and try to use nodes that had your prefix already cached in memory)