LLM Cutover

Guide

A cheaper alternative to GPT for classification

If your GPT calls answer the same narrow question again and again (is this a lead, is this ticket urgent, which category), you are paying general-model prices for a classifier’s job.

What it costs

Per 1,000 yes/no calls at 500 input and 120 output tokens, list prices checked 2026-11-03:

Model$/M in / outPer 1,000 calls
GPT-6.1 Sol$2 / $10$2.200
GPT-6.1 Terra$1 / $5$1.100
GPT-5 mini$0.25 / $2$0.365
GPT-6 Luna$0.1 / $0.5$0.110
Levanto Sage 2$0.05 / $1$0.027

Levanto Sage 2 returns a decision and a probability rather than text, so it bills almost no output tokens. Reasoning tokens on GPT models would add to their rows.

Per-model estimates with reasoning tokens included: GPT-6.1 Sol, GPT-6.1 Terra, GPT-5 mini, all models.

When a small model is a good fit

  • The answer is yes/no, a label from a fixed set, or fields pulled from a document.
  • The same prompt runs thousands of times with different inputs.
  • You can tolerate sending the uncertain share to GPT.

When to keep GPT

  • Open-ended answers, multi-step reasoning or tool use.
  • Low volume: the saving won’t cover the work.
  • Inputs that change a lot from one call to the next.

Check before you switch

Paste 25 real inputs with the labels GPT gave into the free cutover audit. It reports agreement, the confidence threshold that keeps agreement at your target, and the monthly bill before and after. Or estimate first with the routing calculator.

Questions

Is a small model as accurate as GPT for classification?

Sometimes, and only for the task you test. Narrow yes/no and labelling tasks are where small models most often match a large one. Compare answers on your own inputs; public benchmarks won’t tell you about your labels.

What if the cheaper model is unsure?

Send those calls to GPT. With a confidence threshold you keep the cheap answer only where the small model is confident, so quality on the rest is unchanged.