Disclaimer: I use LLMs for support and deep dives into existing code, not for creating new substantial code that I intend to publish.

Quoting from the article:

> With these subsidized token prices currently, I bet we're seeing less and less AI usage encouragement in companies in the future.

I see a different trend, and for a different reason. Employees are encouraged to use open-weight LLMs hosted on-premise, or, at least, in the "correct" geographical area, in preference to models from big providers. The motivation is mostly compliance with GDPR and NDAs. And when this happens, complaints like this get ignored:

> Your local model wrote a wrong root cause analysis for the customer ticket. It's a broken tool, worse than nothing. Claude did it right. We really need either access to frontier models or, alternatively, reasonable expectations on your side about what is doable with the tools you provide.

And that's bad.