> lowering quality of models and other anti users things

Could someone who really follows this closely opine on whether this allegation is really true? Of course I've seen this claim made over and over again to the point that it seems like conventional wisdom. But I also see researchers assert that they haven't actually seen any change in model benchmark scores over time. So I'm still unsure of whether this is a confirmed fact, or an urban legend.

It could even be A/B testing as I have seen asserted and to be fair I have no idea but everything left a bad taste. I don't really understand the API decision this is probably the biggest thing for me and even more important now. People want to use their own harnesses or even just insert llm calls in their scripts, Why pay unsubscribed rates for this?

> I don't really understand the API decision this is probably the biggest thing for me and even more important now. People want to use their own harnesses or even just insert llm calls in their scripts, Why pay unsubscribed rates for this?

I agree with this and it's the single biggest thing driving me away from Claude and towards the OpenAI, Google, and open-weight models. I guess they do this to increase use of their own tools and, thereby, make their models more 'sticky.' But I really dislike it.

it's not, it's two things 1. people getting used to a specific level of capability 2. people exploring capabilities and naturally finding weak spots over time

I expect it to continue, founders hold "Class F" shares, I have a loose guess as to what F stands for.