ran openzoo/auto against openrouter's own auto. same prompts, same minute, 3 runs each.
trivial ("reply with only: ok") openzoo $0.0000054 avg openrouter $0.0000428 avg → 7.9x
code (parse a 2k-line stack trace, group by package, top 5 hot paths) openzoo $0.0000493 avg openrouter $0.0009786 avg → 19.8x
the why: openrouter sent deepseek-v4-flash to all six calls and billed the code one at $0.00097856 three times identically. one model, one price, whatever you asked.
openzoo sent flash to the trivial one, and phi-4 / ling-2.6-flash / solar-pro4 to the code one across three runs. it races candidates and takes whoever lands first.
honest costs: we're slower (10.3s median vs 3.1s on trivial), and because it races, you don't get the same model twice. cheaper and less predictable is a real trade, not a free lunch.
$1k/year buys a lot of calls at $0.00005.