A million labels a month
Classification sends a short text and gets back one word, so you pay almost only for input. The example is 1,000,000 short texts a month, about 500 tokens in (the text, the instructions, and the labels) and 10 out.
| Model | Input | Output | Per month |
|---|---|---|---|
| Claude Sonnet 5.5 | $1,000 | $100 | $1,100 |
| Claude Haiku 5.5 | $50.00 | $5.00 | $55.00 |
| GPT-6.1 Terra | $500 | $50.00 | $550 |
| GPT-6 Luna | $50.00 | $5.00 | $55.00 |
| GPT-5 mini | $125 | $20.00 | $145 |
| GPT-5 nano | $25.00 | $4.00 | $29.00 |
| GPT-4o mini | $75.00 | $6.00 | $81.00 |
| Gemini 3.1 Flash-Lite | $125 | $15.00 | $140 |
| Gemini 2.5 Flash-Lite | $50.00 | $4.00 | $54.00 |
| Grok 4.3 | $625 | $25.00 | $650 |
Lowest here: GPT-5 nano at $29.00, against $1,100 on Sonnet 5.5.
50,000 extracted records a month
Extraction reads a document and writes a small JSON record. The example is 50,000 documents a month, about 2,000 tokens in and 250 tokens of JSON out each.
| Model | Input | Output | Per month |
|---|---|---|---|
| Claude Sonnet 5.5 | $200 | $125 | $325 |
| Claude Haiku 5.5 | $10.00 | $6.25 | $16.25 |
| GPT-6.1 Terra | $100 | $62.50 | $163 |
| GPT-6 Luna | $10.00 | $6.25 | $16.25 |
| GPT-5 mini | $25.00 | $25.00 | $50.00 |
| GPT-5 nano | $5.00 | $5.00 | $10.00 |
| GPT-4o mini | $15.00 | $7.50 | $22.50 |
| Gemini 3.1 Flash-Lite | $25.00 | $18.75 | $43.75 |
| Gemini 2.5 Flash-Lite | $10.00 | $5.00 | $15.00 |
| Grok 4.3 | $125 | $31.25 | $156 |
Lowest here: GPT-5 nano at $10.00. Gemini rows are for comparison; FUSR.ai does not route to Gemini.
Why small models usually win here
Picking a label from a fixed list or copying fields into JSON is a narrow job. Small models do it well once the instructions and the output format are clear. Put the label list and a few examples at the start of the prompt so they come from the cache, ask for structured output, and cap the output length.
Test it on labeled data
You can know for sure in an afternoon:
01Take 200 items you already labeled
Or label them by hand. Include the hard cases.
02Run them through two or three small models
Same prompt, same output format.
03Count the mistakes
Pick the cheapest model whose mistake count you can live with, and send only the unsure cases to a larger one.
Sending the unsure cases to a larger model is routing. You can build it yourself or use an AI API router.