How a support answer is priced by the model provider
Every AI provider bills the same way: a rate per million tokens that go in (your prompt) and a rate per million tokens that come out (the answer). A support reply is a small amount of output and a surprisingly large amount of input, because a useful answer has to read your help center first.
The shape we use below is a real one: the customer's email plus the retrieved help-center articles, about 25,000 tokens in, and a short reply of about 500 tokens out. A tiny help center makes it cheaper; a sprawling one makes it dearer. Change the assumption and the ranking does not move.
The table
Per answer and per month at 1,500 AI-handled conversations. Rates as published on 2026-07-27; cheapest first.
| Model | Per answer | Per month |
|---|---|---|
| DeepSeek V4 Flash (DeepSeek) | $0.004 | $5 |
| GPT-OSS 120B (Groq) | $0.004 | $6 |
| Mistral Small (Mistral) | $0.004 | $6 |
| GPT-5.4 nano (OpenAI) | $0.006 | $8 |
| Gemini 2.5 Flash (Google) | $0.009 | $13 |
| DeepSeek V4 Pro (DeepSeek) | $0.011 | $17 |
| Llama 3.3 70B (Groq) | $0.015 | $23 |
| GPT-5.4 mini (OpenAI) | $0.021 | $32 |
| Claude Haiku 4.5 (Anthropic) | $0.028 | $41 |
| Grok 4.3 (xAI) | $0.033 | $49 |
| Mistral Medium (Mistral) | $0.041 | $62 |
| Grok 4.5 (xAI) | $0.053 | $80 |
| Claude Sonnet 5 (Anthropic) | $0.083 | $124 |
| Claude Opus 5 (Anthropic) | $0.138 | $206 |
Now the same column for a metered helpdesk
Intercom prices Fin at $0.99 per outcome, on top of seats. At 1,500 outcomes that is $1,485 a month. Help Scout's AI Answers add-on is $0.75 per resolution: $1,125 a month. The cheapest row in the table above is $5, and the most capable models are still a fraction of either.
That gap is not the AI. It is the meter. The answer is the same answer; one bill has a markup on it and one does not. And because the meter runs per outcome, it grows precisely when the AI does its job well.
What you can do with this
- Ask your vendor what you paid per AI outcome last month. If the number is not on the invoice, that is an answer too.
- Put it next to the table. Most teams land between $0.004 and a few cents per answer on their own key.
- If the gap matters, run the AI on a key you own. That is what Kerux is: a helpdesk where the AI bill goes to your provider, at cost, and we only charge per agent.
Competitor figures are from the vendors' pricing pages, verified 2026-08-07. Model rates are the providers' published list prices on 2026-07-27; they change, so check before you quote them.