Skip to the document
Madhuopen lab

AI Insights: Latency & Cost Optimization · part 2 of 17

rejected23 July 2026cost: included in baseline phase

Cheaper paid models: can a smaller GPT do the job?

GPT-5-mini, GPT-5-nano and older mid-tier paid models on the same task

The question

Before touching open models, can a cheaper paid tier match GPT-5.1 on this task?

rejected

Nothing cheaper on the paid side matches GPT-5.1. The mini/nano reasoning tiers add cost and delay on a short structured task.

GPT-5-mini first word22.7 sthinks before answering
GPT-5-nano first word59.8 s64 s total
Cost per answer$0.022–0.029MORE than GPT-5.1
Qualitynot bettermore text, same accuracy

Results

Results
ModelFirst wordTotal timeCost/answerNote
GPT-5.1 (reference)1.5 s24 s$0.0136quality bar 8.7
GPT-5-mini22.7 s34.6 s$0.022reasoning overhead, no quality gain
GPT-5-nano59.8 s64.0 s$0.029worst first-word time measured
Older mid-tier paid modelscheapervisibly weaker on tax content

Why the small tiers fail here

The mini and nano tiers reason step by step before answering. For a short, structured extraction-and-advice task, that thinking phase means 8–60 seconds of dead air before the first word — the exact thing this project is trying to fix — and the extra generated text makes them cost more per answer than the bigger model they were supposed to undercut.

GPT-5-miniGPT-5-nanoreasoning models