Trying to get my AI spending under control, I’m having the server randomly switch to the “flex” pricing tier in GPT-5 when there’s background work that doesn’t need to be instantly fast. So far so good.

@lukebouch Still fairly limited in Micro.blog. Mostly things like accessibility text in photos and transcripts for podcasts. The big new thing is Inkwell’s “Reading Recap” feature, which creates a personal recap of last week’s blog posts in your subscriptions. That bumped up usage by over 10x.

@manton I've built a local "app" (nearly an agent) that determines which models to run based on prompt heuristics. I have several local models, an image model, and several cloud based models tuned so that it knows which to call when.

It works pretty well. I’m absolutely positive something better will be created by someone else (if it doesn't exist already) but I'm still thinking of open sourcing it.