We benchmarked Ox Alpha vs Fable on a real bug from the Cline repo. Both fixed it correctly.

But we found that Ox used much fewer thinking tokens. Most reasoning models loop and re-derive the same conclusion over and over before acting (Fable said "I found the root cause" 7 times before editing).

Ox stated it once then wrote the fix. Roughly ~3x lower output tokens for the same work.

Reasoning models have been trained to increase reliability with re-verification. Ox seems to trust its first conclusion instead, which feels like a fundamentally different post-training philosophy.