How Much Does an AI Chatbot Cost in 2026? Real Numbers by Scale
2026-07-09
"How much does an AI chatbot cost?" has two honest answers: the token bill (surprisingly small at low scale) and the total cost (where most budgets break). This guide computes both, at three scales, using prices verified on 2026-07-05 — so you can budget before you build. For your own numbers, use the free calculator.
The three cost layers of every chatbot
- Tokens — what the model provider charges per request. Scales with users × questions × response length.
- Infrastructure — hosting, and if the bot answers from your docs (RAG): embeddings + a vector database. Typically $50–300/month for small-to-mid deployments.
- Humans — reviewing transcripts, handling escalations, maintaining prompts. The layer everyone forgets and the biggest one at scale.
Token bill by scale (computed, not guessed)
Assumptions per request: ~400 input tokens (system prompt + question + a little history) and ~300 output tokens — a typical support exchange. All three scales below use them.
| Scale | Volume | Economy model | Mid-tier model |
|---|---|---|---|
| Small (100 users × 5 q/day) | 15k req/mo | Gemini 2.5 Flash: ~$13/mo | Claude Haiku 4.5: ~$29/mo |
| Medium (500 users × 10 q/day) | 150k req/mo | GPT-5.4 mini: ~$248/mo | Claude Sonnet 5: ~$570/mo |
| Large (5,000 users × 10 q/day) | 1.5M req/mo | Gemini 2.5 Flash: ~$1,305/mo | GPT-5.4 mini: ~$2,475/mo |
Two lessons hide in that table. First: at small scale, tokens are almost free — your real costs are infrastructure and your own time. Second: at large scale, model choice is a multi-thousand-dollar decision — see our mini-tier comparison for how the cheap tiers stack up.
Caching: the discount most chatbots leave on the table
Every request repeats your system prompt and knowledge snippets. Providers bill that repeated prefix at ~10% of the input rate once you enable prompt caching. On the medium scenario with Claude Sonnet 5, a 60% cache-hit rate cuts the bill from ~$570 to ~$505/month — and input-heavy bots with long knowledge contexts save far more.
The costs that sink chatbot budgets
- Context growth — every conversation turn resends history; long chats cost multiples of short ones. Trim or summarize history.
- RAG infrastructure — a docs-answering bot needs embeddings + a vector database: budget $70–200/month managed.
- Escalation handling — if 10% of 150k conversations escalate to humans at even 3 minutes each, that's 750 hours/month of payroll. Design the bot to deflect well, not just answer.
- Retries and abuse — rate-limit your endpoint; bots and abuse traffic burn real tokens.
So what should you budget?
A sane starting budget for a production support chatbot: token bill × 1.3 (safety margin) + $100–300 infrastructure, plus a defined human review workflow. For the medium scenario on a mini-tier model, that's roughly $420–620/month all-in — usually a fraction of one support hire. Run your exact assumptions in the calculator (it models every layer above, including break-even against what you charge), and if you want an expert sanity-check, request a free AI cost review.
Prices verified 2026-07-05 on each provider's official page. Estimates are for planning purposes only — always confirm current pricing with the provider.