Tempted by the promise of 95% cheaper tokens, I decided to test out Hermes + Minimax3 via Fireworks this weekend on a personal project, with Codex + gpt5.5-xhigh as my benchmark.
The job was to build me a custom viewer for my home security cameras, which come with some truly horrendous software. I provided NVR credentials and no other help.
Codex finished the job in 2 shots, spending 100k tokens. Mini spent 180M tokens over more than 20 shots. 1800x as many tokens made Mini cost 90x as much despite the per token price being 1/20th.
Just one anecdote, but looks like power law returns to intelligence are still in place for some things.