I must say it works beautifully now! I'm still torn between using the default config or premium but default might just be good enough and the simple things get really cheap.



Sometimes I wish it would have more aggressive thresholds just to be save it's not missing nuance but that's probably fine tuning and not worth another intermediate config.

I have long sessions and the queries are surprisingly cheap. As I told you, the switch to "save a bit" resulted in record high queries consistently.

These are all from the same session and the 83ct one is where I switched to autoclaw/premium.

But ... it might be false alarm. Sorry. It would be great if the activity list would show if it was autoclaw-kimi or just kimi but just now I'm getting kimi and before might have been opus being configured on the specific agent, ignoring the default model that I had changed to autoclaw. Sorry. Will test again and please leave autoclaw/premium in place! 🙏

🙏 love your service. reverted back for now but would really love if your service could actually be a bit cheaper automatically depending on content as I'm using it more and more.

Given the financial reality, it's bound to happen.

Cost to do so for somebody who knows to hide their tracks: Virtually zero.

Potential gain from shorting the market: More than zero.

And then there is those that gain from Bitcoin losing, so ... I wish somebody had done that 10 years ago but at some point it will happen. Even if Knotzis get their way cause data continuity is the biggest joke in their argumentation.

For debugging: I did use autoclaw/premium but paying significantly more for the same opus suggests some cache issues or something. 👀

Wow, well, ... guess I know what not to use until you announce some fix for whatever that is.

Seriously @PayPerQ I need you to look into this. I checked my account activity. Most is opus-4.6. None of the last 200 api calls is over 61ct. Except for "What is 5+5". What gives? Was this shifting cache from some place to another? I'm scared to run a test with more calls.

So they are spying but can you elaborate? Who might be using these models without knowing it?

I'm wondering how this is supposed to work, given the context is not only my last input to openClaw or will that trigger once the sub-agents get smaller tasks? Actually my test prompt about 5+5 was the most expensive I asked an LLM in a long time. Isn't the idea that the 15 dimensional chess would lead to the conclusion that for the final prompt, the prior conversation context is not needed, thus a cheap model can handle it?

Not sure. I'll ask him when I get a chance.