We’ve been using gpt-5.6-luna lately in our production systems. I finally figured out that the pricing that is quoted is misleading. The claim is that the price is $0.20 per million input tokens. You pay 1.25x for write cache and 0.1x for read cache. You can choose explicit mode or implicit mode for caching. In explicit mode, you tag each input block that you want to be written to cache. Other inputs won’t be cached. In implicit mode, the system takes care of it for you.

At least for us the main reason that this pricing is misleading is that, in practice, there is no real option NOT to use implicit mode. And that means that ALL of your input tokens are getting charged at 25% above that stated price (except for the cache hits of course).

So you might say, “Well, just use explicit mode. You decide what to cache.” Yes, but there’s a big catch. When you use function tools (as we do extensively), each tool call generates tool results that have to go back to the model. And while implicit caching will work on those, there’s no way to mark these tool results as cacheable. Normally you wouldn’t think that a tool result would be cacheable. But remember that when making a chain of calls, the recommended process is to append the tool results to the end of the same input chain. So all of your stuff that didn’t appear to be cacheable, including those tool results, really needs to be cacheable for the subsequent steps in the same response generation (for each set of tool calls made by the model).

Note that this isn’t really about the tool results at all. Let’s say that you have 10k input tokens that are cacheable and another 25k that are not cacheable (different on each generation). But if you call a tool, you’ll be calling the model again with the tool result, and you need to include those 35k original tokens again – so you better mark those 25k as cacheable even though they didn’t seem cacheable to you!

In our case, the tool results are sizeable as well, and there isn’t even a way to mark them as cacheable (although it appears that in implicit mode, OpenAI will cache them for subsequent cycles).

So if you use any tools at all, forget about using explicit mode. And in implicit mode, you’re really just paying 1.25x for all input tokens for which there is no cache hit.

For me, the whole thing would have been more transparent if you had simply made the price of Luna $0.25 per million input tokens and $0.02 per million cached input tokens.

Don’t get me wrong. I appreciate the low price on Luna. It is the best value in the industry right now. But I just don’t like the bait-and-switch on the way the pricing is described. Seems a bit slimy.