← Guides · November 8, 2026
How Much Does an AI Chatbot Cost to Run? GPT vs Claude vs Gemini
Most teams guess their AI bill and discover the real number on the first invoice. You can do better in five minutes, because chatbot cost follows a simple formula:
monthly cost = requests per month × (input tokens × input price + output tokens × output price)
Prices are quoted per million tokens, so divide by 1,000,000 at the end.
A realistic support chatbot
Let’s model a mid-size support bot:
| Assumption | Value |
|---|---|
| Conversations per day | 1,000 |
| Messages per conversation | 5 |
| Requests per day | 5,000 |
| Input tokens per request (system prompt + history + question) | 1,200 |
| Output tokens per request | 350 |
That is 150,000 requests and about 232 million tokens per month. Here is what that costs at list price (November 2026), without any caching:
| Model | Tier | Monthly cost |
|---|---|---|
| GPT-6 Luna | budget | ≈ $44 |
| Claude Haiku 5.5 | budget | ≈ $44 |
| Gemini 3.8 Flash | budget | ≈ $332 |
| Claude Sonnet 5.5 | workhorse | ≈ $885 |
| GPT-6.1 Sol | workhorse | ≈ $885 |
| Gemini 4 Argon | workhorse | ≈ $885 |
| Claude Opus 5.5 | flagship | ≈ $1,770 |
| GPT-6 Astra | flagship | ≈ $4,425 |
The spread is 100× between the cheapest and most expensive option for exactly the same traffic. Model choice is the single biggest decision.
Lever 1: pick the smallest model that passes your tests
Build a test set of 50–100 real questions, run them through a budget model and a workhorse model, and compare answers. Many support bots work fine on a budget tier for routine questions and only escalate hard ones to a bigger model. A 90/10 split between Haiku-class and Sonnet-class models costs roughly a seventh of running everything on the larger one.
Lever 2: prompt caching
Your system prompt, policies and tool definitions are identical in every request. With prompt caching, those repeated tokens are billed at the cache-read rate, which is 90–97% cheaper than fresh input. If 40% of each request’s input is cacheable, the workhorse bill above drops from about $885 to roughly $750 per month.
Lever 3: trim the output
Output tokens cost five times more than input tokens on most models. Asking the model to answer in at most three sentences, or returning structured JSON instead of prose, often halves output length without hurting quality.
Run your own numbers
Every app is different. Open the calculator, pick the Chatbot preset and adjust the traffic, token counts and cache share to match your logs. The results update instantly and the price table shows the exact rates used.