We burned 1.5B OpenAI tokens — here's the postmortem that cut our API bill 65%

Careful-Key-1958 · reddit · 2026-09-19

Borut, founder of Rankpilot.dev, shares a postmortem after his team processed 1.5 billion OpenAI tokens in a month, ending with 65% lower monthly costs, faster responses, and better output quality.

Measure before optimizing: everything ran through one API key, so they couldn't see which feature was burning money; they now tag every call by feature and log tokens/cost per feature.
Half the "AI" calls didn't need AI: regex/parser checks (e.g., title tag length) run first; only fuzzy questions like tone or intent matching go to the model, removing 5% of calls.
Model cascade: a small model answers first with a confidence flag; only 5% of requests escalate to the expensive model.
Content hashing: pages are re-analyzed only when their cleaned content hash changes — one of the biggest wins.
Trim inputs: strip nav bars, cookie banners, footers and scripts; keep only main content with a size cap.
Small eval set: 20 hand-checked examples run before every model/prompt change; twice this caught cheaper setups that quietly degraded.
Fix double billing: timeout retries were resending full requests while originals kept billing; idempotency keys, exponential backoff, and retry caps fixed it.

Original post →

We burned 1.5B OpenAI tokens — here's the postmortem that cut our API bill 65%

More from coding & agent