Ten thousand OpenAI API requests cost about $6 on GPT-4o mini and $100 on GPT-4o, assuming a typical 2,000-token prompt and a 500-token reply. The model you pick from habit matters more than almost anything else in your AI budget.
My opinion: start every store workload on the cheap model and make the expensive one prove itself. Most of the bloated AI bills I see come from defaulting to a frontier model for work a small model handles fine, and from oversized prompts. The API price itself is rarely the issue.
The GPT vs Claude vs Gemini cost calculator shows the same comparison for your own token sizes.
The math in one line
Cost per request = input tokens × input rate + output tokens × output rate. On GPT-4o mini ($0.15 / $0.60 per million tokens):
(2,000 × $0.15 + 500 × $0.60) ÷ 1,000,000 = $0.0006 per request → $6 per 10,000.
Same 10,000 requests, four models
| Model | Rate in / out (per MTok) | Per 10k requests |
|---|---|---|
| GPT-4o mini | $0.15 / $0.60 | $6 |
| Gemini 2.5 Flash | $0.30 / $2.50 | $18.50 |
| Claude Haiku 4.5 | $1 / $5 | $45 |
| GPT-4o | $2.50 / $10 | $100 |
That is a 16× spread for the same prompt.
Worked examples at different scales
- One-off batch (say, rewriting 10,000 product descriptions): $6–$100. At this size, choose for quality. The difference is lunch money.
- A store assistant at 1,000 requests a day (30,000/month): $18 on mini against $300 on GPT-4o.
- A busy support bot at 10,000 requests a day (300,000/month): $180 on mini against $3,000 on GPT-4o. At this point the gap pays for a proper routing setup.
- Bloated context (6,000 tokens in, 800 out, still on mini): $13.80 per 10k. You’ve more than doubled the bill without changing models.
When the cheap model isn’t enough
Move hard requests to a bigger model, and only those, when:
- Multi-step tool use (order lookup → stock → shipping exception) keeps losing the thread
- Refund or policy edge cases fail your test set often enough to cost more than the price difference
- A wrong answer is costly: wholesale, regulated products, VIP customers
If 10–20% of traffic needs the big model, route that 10–20%. Upgrading 100% of traffic is how a $180 bill becomes $3,000.
What I’d do
- Measure your real token sizes. Log a day of requests and average the input and output tokens. Most teams guess low on input because of system prompts and retrieved context.
- Build a 50-example test set from real traffic and run it on mini first. If mini passes, ship it. If it fails on specific intents, route only those intents up.
- Trim prompts before you upgrade the model. Use the top 3–5 retrieved chunks, not the whole help PDF, and a short system prompt. Then put the trimmed numbers into the calculator to set your monthly budget.
For a broader walkthrough of estimating OpenAI costs, see how to estimate OpenAI API costs.
Work with me
I’m Faraz, a Shopify and AI automation consultant. I build support bots, catalog pipelines and model routing for stores, sized so the AI bill stays predictable. farazhussain.com · WhatsApp +92 334 3454747.
Prices are approximate 2026 list rates. Check current pricing on OpenAI, Anthropic and Google before you budget.