Comment by ravenstine

> I have no idea what type of job mini can do that nano couldn't or vis-à-vis.

In my experience, Nano won't reliably handle complex open-ended tasks and is mostly suited for very explicit instruction that it can't screw up. It's no different from how there are some chores you can give to kids and there are other tasks you need at least a teenager for. If the decision tree of the task is very clear and conventional, Nano can be cheaper than giving the task to a relatively overpowered model, especially if it's something where the output is rigidly structured. This makes it well suited for skills that essentially run CLI commands and generate output, especially because it is usually faster. Mini is more like a discount version of the base model, and Nano is the dollar store version. Mini is more of a generalist and a fairly good deal if you have a moderately complex task that is conventional, but can be less conventional that what Nano can handle. I mostly used gpt-5.4-mini this year for my side projects because it's a pretty good generalist while significantly saving on costs. It is, however, somewhat dumber than the base model and more prone to ignore or forget rules you give it. I'd have just used a base model, but the low cost of Mini and Nano made them appealing to me. Maybe I'm a cheapskate, but I have hundreds or possibly thousands more in my pocket than many other users because of that.

This workflow I settled into with Mini and Nano didn't map cleanly on to the current generation of model tiers. With the price of Luna, you'd think it would be a replacement for Nano. In a sense it is, yet I didn't find that Terra became the new Mini. Terra is more powerful, better at explaining its own decisions, yet I've also found it to be relatively stupid while charging me more to use it. On the other hand, Luna with its reasoning set to "high" is what I consider to fill the role of Mini, and is good enough such that I no longer use Mini. Sol and Astra are great, but they're pricey. It could be my own brain and its bad perception, but so far I don't get the point of Terra. Luna succeeded at reverse engineering some abandonware with a very complicated licensing and virtualization scheme, and did so over SSH into a Windows VM with only PowerShell on the other end. Terra did such idiotic crap to my flashcards app that I stopped using it for anything after that.

This is why I find OpenAI's naming unhelpful and kind of pointless. I don't really care about the benchmarks that all these models are commonly run against. They're not that useful, IMO. OpenAI could easily give early access to these models, get a ton of feedback, and provide better insight to customers on how these things behave. Even calling Terra "gpt-5.6-overpriced-cheating-dumbass" would be better than wasting my time and money figuring it out myself. But that wouldn't make OpenAI as much money.