What Labeling 10,000 Emails Costs on Three LLM APIs, Step by Step
Labeling 10,000 emails with a short instruction prompt costs $0.85 on GPT-6 Luna, $0.85 on Claude Haiku 5.5 and $17.00 on GPT-6 Sol. The two small models land on the same number because their rates match to the cent. The twenty-fold gap to Sol is the whole story of this estimate, and everything below is the arithmetic behind it.
Rates are as of November 6, 2026, per million tokens, from the OpenAI pricing page and Anthropic's Claude pricing page. The token counts are assumptions made for the example. Replace them with yours.
The job, in numbers
Each request carries a fixed instruction of 400 tokens (the label list and rules) and one email of 350 tokens on average. The model answers with a label and a confidence, about 20 tokens. There are 10,000 emails.
- Input per request: 400 + 350 = 750 tokens, so 7.5M input tokens in total.
- Output per request: 20 tokens, so 0.2M output tokens in total.
- Input cost = 7.5 x input rate. Output cost = 0.2 x output rate.
Three models, three bills
Rates used: GPT-6 Luna $0.10 in and $0.50 out; Claude Haiku 5.5 (prompts up to 100K tokens) $0.10 and $0.50; GPT-6 Sol $2.00 and $10.00.
| Model | Input | Output | Total | Instruction cached | Batch |
|---|---|---|---|---|---|
| GPT-6 Luna | $0.75 | $0.10 | $0.85 | $0.49 | $0.425 |
| Claude Haiku 5.5 | $0.75 | $0.10 | $0.85 | $0.49 | see below |
| GPT-6 Sol | $15.00 | $2.00 | $17.00 | $9.80 | $8.50 |
Source: OpenAI API pricing and Claude pricing, as of November 6, 2026; token counts are the assumptions above.
How the other columns come out. Caching the 400-token instruction means 4M cached tokens and 3.5M fresh ones. On Luna the cached rate is $0.01, so 4M x $0.01 + 3.5M x $0.10 + $0.10 of output = $0.49. On Sol the cached rate is $0.20: $0.80 + $7.00 + $2.00 = $9.80. Haiku 5.5's cache-read rate is also $0.01, which gives the same $0.49, but that leaves out the write charge ($0.125 per million) and Anthropic's page says cache pricing reflects a 5-minute TTL, so a slow queue pays for writes more than once.
OpenAI's batch rates for these models are half of standard: Luna $0.05 in and $0.25 out, Sol $1.00 and $5.00. That gives 0.375 + 0.05 = $0.425 and 7.5 + 1.0 = $8.50. Anthropic's page says "Save 50% with batch processing" but lists no batch table, so the Haiku cell stays open rather than halving a number the page does not print.
What moves the price most
Model choice, by a wide margin: $0.85 against $17.00 for the same tokens. Cache comes second on Sol, cutting $17.00 to $9.80, about 42%.
The surprise is output length. A label is 20 tokens. If a model starts explaining itself and writes 200 tokens per email, output grows to 2M tokens. Luna rises to $0.75 + $1.00 = $1.75. Sol rises to $15.00 + $20.00 = $35.00, more than double, because output is billed at five times the input rate on both OpenAI models here.
One more caveat on Haiku. The site's Haiku 5.5 price note covers the updated tokenizer, which takes slightly more tokens for the same text and has no published multiplier. If emails came out 20% longer in tokens (9.0M in, 0.24M out), the bill would be $0.90 + $0.12 = $1.02 instead of $0.85. That is an illustration, not a measured figure.
Checking your own numbers
Paste a typical email and your instruction into the calculator on the home page to get token counts. It shows April 2023 rates and counts with OpenAI-style tokenizers, so use it for rough counts only (Anthropic's tokenizer counts differently) and take the prices from the two pages above. Then replace 350, 400 and 20 in the steps and rerun the table. Prices change, and those pages win over this one.