Skip to main content

​
The scenario

Imagine a support team using GPT-4 to automatically route customer emails to the right department. Technical issues go to engineering, billing questions head to finance, and sales inquiries land with the sales team. It’s a straightforward task that works well, except for one problem: the costs add up fast. The question worth asking is: do you really need the most expensive model on the market just to sort emails into categories?
This guide uses the Narev Pricing API to compare token rates and estimate per-request cost. No SDK required.

​
The hypothesis

Email classification has clear categories and predictable patterns. A cheaper model might handle the task just as well. Before you switch in production, pull live rates and price a representative request.

​
Compare token rates

Search pricing for candidate models, then read pricing.prompt and pricing.completion (USD per token):

​
Variants tested

​
Price a representative request

Send token counts for a typical classification call. The API returns a dollar total:
Repeat for each candidate model using the same usage block. That isolates rate differences from prompt or output length changes.

​
What to track

  • Cost per request from cost_breakdown.total
  • Latency in your own app (the Pricing API does not measure this)
  • Accuracy on a held-out set of emails (run offline or in staging)

​
Results

GPT-4 is the expensive baseline: 100% accurate in this test set but costly. gpt-oss-20b cuts cost by 99% while keeping 100% accuracy in the original experiment. The ultra-cheap nano variants save more on paper but dropped accuracy too far for production.
Model experiment results comparison

​
Winner: gpt-oss-20b

For this routing workload, gpt-oss-20b matches GPT-4 accuracy at a fraction of the cost. Latency is about one second slower per request in the original run. That tradeoff is often acceptable for async email routing.

​
Example in action

A typical sales inquiry should route to the sales team:
Subject: Enterprise plan pricing
From: cto@fastgrowth.ioHi there,A 200-person company is looking to upgrade. Could you send details about:
  • Enterprise plan features
  • Volume discounts
  • Implementation timeline
  • API rate limits
Please schedule a demo for next week.Thanks,
Mike Chen, CTO
Should route to: Sales team

​
How each model did

Individual model result comparison
Every model routed this email correctly. Edge cases matter more than easy wins when you pick a cheaper model.

​
The takeaway

Task complexity should drive model choice, not defaults. For email classification, gpt-oss-20b is the sweet spot: about 99% cheaper than GPT-4 with 100% accuracy on the test set.
Overall impact of model optimization
Use GET /v1/prices/search to find candidates and POST /v1/traces/cost to price your workload before you cut over.
Need live rates in your app? Start with the quickstart or the API reference.