Dev finds GPT-OSS-120B ran at medium reasoning effort for months; switching to high fixed quality

userpostingcontent · reddit · 2026-09-24

A solo developer behind PopUpFactCheck shares a "quality breakthrough": for months, the primary GPT-OSS-120B model had been running at OpenRouter's medium reasoning effort instead of high. Switching to high substantially improved the model on the hardest attribution, evidence reconciliation, and judgment tasks.

Because the architecture aggressively caches completed fact-checks in FAISS and DynamoDB and routes inference to the cheapest provider, the extra reasoning cost is absorbable. The change is live in production.

Lesson: before tuning prompts, check low-level settings like reasoning effort.

Original post →

More from coding & agent

coding & agent channel →