site banner

Friday Fun Thread for August 7, 2026

Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.

1
Jump in the discussion.

No email address required.

they're making unlimited usage of a newer model, 5.6 Luna, available to the masses for free.

In my ime, all cost optimized models are absolute worthless dogshit. A current gen cost optimized model is not as useful as early gen normal models (llama3 70b, gpt-4o).

My understanding is that current chatgpt free is 5.5 real with thinking=0 which is still a decent model. Luna (equivalent to nano, much worse than even mini), is a maaaasive step down. I think it's criminal to even let the masses have access to this piece of hot garbage, as it will provide negative information to a 100iq midwit.

This is true and I'm currently doing a project that has ABSURD unavoidable token read/write overhead (I'm processing ~300 cookbooks) and it takes 100 millions of tokens per book to ensure a reasonable level of accuracy.

So I'm basically stuck with subscription subsidized tokens, but Codex only serves the most recent models. I'd absolutely kill for a gpt-5 full size at low cost, but they don't offer the big old models at low cost (presumably bc expensive to run, but they're still much smaller than the base model of 5.5+) and Luna, while now gloriously cheap, is just so fucking stupid and myopic.

I could do this project with a 2 year old LLM most likely, it's just so deeply unergononic

Luna is incredibly useful if you have a task scoped right for it. Funnily enough, low effort seems to perform better than high/xhigh, because the latter keep trying to find a way to galaxy-brain their way around simple evaluation tasks.