[captainmuon]

It's one thing that the APIs don't return the reasoning tokens in a readable fashion. But then why do I have to send them the whole conversation every time? If they don't allow me to see parts of the conversation then they should have the burden of storing and transmitting it, and shouldn't charge more and more the longer the conversation gets (unless opaque token caching kicks in) like some LLM Shlemil the painter.

[the_mitsuhiko]

Because if you don't, you trash the KV cache and things get way more expensive. We wrote about it here: https://earendil.com/posts/prompt-caching/

Encrypted reasoning content is one of the things we keep as blobs in the transcript, even though we can only send it back to the original provider that created them, as otherwise we cannot keep caches live.

[captainmuon]

Right, but my point was, I should be able to just send something like

    {'continue_conversation_id': 123, 'message': 'another message?'}

Seems like it would be even less work for the LLM host to just check if ID 123 is still in a cache (and if not, load it from a database) than decoding my request and checking cryptographic signatures. Right now the whole conversation history with blobs functions like a really long ID.

[the_mitsuhiko]

But then on a cache miss you need to resend the whole thing. Cached websocket connections on OpenAI work like this in a way. But once you lose the connection then you lose that.

Your idea though is the core of the responses API with store where everything is stored on their side and then you append to it.

For the LLM there is no difference btw. In either way you need to locate the cache and use it.

[charcircuit]

>and shouldn't charge more and more the longer the conversation gets

The cost of generating token N is O(N) with KV cache so it's unrealistic to expect to not be charged more the longer the conversation is if you are looking for the minimum price.

[captainmuon]

Except, if some of the answered tokens are opaque to me, but I pay for them, the operator has no incentive to keep their number low [1]. For all I know the model could be spinning thumbs while thinking. If they cannot provide transparency, I would prefer that they not charge me for that part. Of course, they would prefer to charge me, so sure it is unrealistic to expect that they don't. I'm not trying to predict what they are going to do as a market participant, just stating my moral preference :-).

Also, isn't it without caching even something like O(N^2) because you have to replay the whole conversation on every request to reach the same internal state? My point was I shouldn't have to pay for cache misses if hitting the cache is not deterministic. Give me a guarantee that you keep the session "hot" for N minutes, and cached on disk for M months, charge a little bit more on average, but then the pricing is at least transparent.

[1] Except the general market pressure to keep total cost for the same problem solved lower than the competition.

[charcircuit]

Ultimately you are paying for quality so you end up relying on [1] regardless. Even if they show you tokens they can cheat by using weaker models, showing fake thinking tokens, etc.

>because you have to replay the whole conversation on every request to reach the same internal state?

It's because you have to rebuild what would have been cached for every token before the latest one that is being worked on.