I still think it's basically unavoidable that most people who might pay for api access will end up on-prem.

Fixed costs, exact model pinning, outage resistant, enshittification resistant, better security, better privacy, etc...

There are just so many compelling reasons to be on-prem instead of dependent on a 3rd party hoovering up all your data and prompts and selling you overpriced tokens (which eventually they MUST be, because these companies have to make a profit at some point).

If the only counterbalance is "well the api is cheaper than buying my own hardware"...

That's a short term problem. Hardware costs are going to drop over time, and capabilities are going to continue improving. It's already pretty insane how good of a model I can run on two old RTX-3090s locally.

Is it as good as modern claude? No. Is it as good as claude was 18 months ago? Yes.

Give it a decade to see companies really push into the "diminishing returns" of scaling and new models... combined with new hardware built with these workloads in mind... and I think on-prem is the pretty clear winner.

I use Gemini-CLI at work, and haven't noticed anything. I use Google Jules (free tier) on a toy project much more heavily and can't complain. I think sometimes the prompts take longer than they used to, but I couldn't care less. I'm not in a hurry.

It's still not uncommon for it to escape it's thinking block accidentally and be unable to end it's response, or for it to call the same tool repeatedly. I've watched it burn 50 million tokens in a loop before killing the chat.

No. It's still shit. It can do some well contained tasks, but it is very less usable on production codebases than gpt or claude models. Mainly because of the usage limits and the lack of good environments for us to use it on. Anthropic gets away with this because claude code, as bad as it is, is still quite functional. Gemini cli and antigravity are utter trash in comparison.