What is the cost per customer support chat using GPT-4o?

A typical customer support chat of 8 messages with growing conversation history costs approximately $0.101 on GPT-4o. At 100,000 chats/month, that totals $10,100/month or $121,200/year.

Why does conversation history make AI chatbots expensive?

Most LLM APIs charge per token in the input — and for multi-turn conversations, every message includes the full history. Message 1 might have 400 tokens; message 8 has 400 + 7 prior exchanges ≈ 2,800 tokens. The cost per message grows linearly with conversation length, making long chats 5-8× more expensive than short ones.

How do I reduce AI customer support costs?

Three main levers: (1) use a smaller model tier for simple queries and escalate to frontier only when needed; (2) implement conversation summarization — compress old history into ~200 tokens instead of keeping the full transcript; (3) cache prompt prefixes — the system prompt and FAQ context can be cached, cutting repeated input cost by 50-90% depending on provider.

What 100,000 customer support chats cost with different LLMs

Q: How much does an AI customer support chatbot cost per month?

It depends heavily on the model and conversation length. At 100,000 chats/month with 8 messages average, Gemini 2.0 Flash costs around $432/month, GPT-4o mini costs $648/month, Claude Haiku costs about $3,744/month, and GPT-4o costs approximately $10,125/month.

Q: Is Gemini Flash good enough for customer support?

Gemini 2.0 Flash costs about $0.0043 per 8-message chat — 23× less than GPT-4o. For many support use cases (FAQ answering, order status, return policy) Flash quality passes A/B tests against larger models. For complex troubleshooting or nuanced tone requirements, benchmark before committing.

Published 15 July 2026 · reference prices, verify before budgeting

"How much will our AI support chatbot cost?" is one of the most common questions we see. The answer varies by 23× depending on the model — and grows faster than you expect as conversation history accumulates.

The assumptions

100,000 chats/month (~3,333/day)
8 messages per chat average (4 user + 4 assistant)
System prompt: 400 tokens (included in every message)
Growing context: message 1 = 400+150 tokens, message 8 = 400+1,250 tokens (history builds)
Average input per message: ~1,350 tokens (weighted mean across 8 messages)
Average output per message: 280 tokens (typical support reply)
Token totals per chat: ~10,800 input + 2,240 output = 13,040 total tokens

These are realistic averages. Your actual token use depends on system prompt length, FAQ context, escalation policy, and how often you summarize history.

Cost per chat and monthly total

Model	$/1M in	$/1M out	Cost per chat	100k chats/mo	Annual
Gemini 2.0 Flash	$0.10	$0.40	$0.0043	$432	$5,184
GPT-4o mini	$0.15	$0.60	$0.0065	$648	$7,776
Claude Haiku 4.5	$0.80	$4.00	$0.0374	$3,744	$44,928
GPT-4o	$2.50	$10.00	$0.1013	$10,125	$121,500
Claude Sonnet 4	$3.00	$15.00	$0.1258	$12,576	$150,912

Gemini Flash/month

$432

GPT-4o/month

$10,125

Difference

23×

Flash annual saving vs GPT-4o

$116k

The conversation history effect

This is where most chatbot cost estimates go wrong. They model cost per message at a fixed token count — but in reality, each message carries the full history.

Message #	Input tokens (incl. history)	Output tokens	Flash cost	GPT-4o cost
1	550	280	$0.000167	$0.00418
2	900	280	$0.000202	$0.00505
4	1,400	280	$0.000252	$0.00630
6	1,950	280	$0.000307	$0.00768
8 (final)	2,500	280	$0.000362	$0.00905
Total (8 msg)	~10,800	~2,240	$0.0043	$0.101

Message 8 costs 2.2× more than message 1 on any model — purely from context accumulation. For long support flows (password reset + account lookup + order history = 15+ messages), costs grow further. A 15-message chat on GPT-4o costs roughly $0.22 per conversation.

Three levers to cut costs

1. Model routing (biggest lever)

Not all support queries need the same model. Route queries to different tiers:

FAQ / simple queries (60% of volume): Gemini Flash — $432/month for 60k chats
Medium complexity (30%): GPT-4o mini — $194/month for 30k chats
Complex / escalated (10%): GPT-4o — $1,013/month for 10k chats
Total blended: ~$1,639/month vs $10,125/month for all-GPT-4o

A simple intent classifier (itself cheap — Flash at ~$0.002 per routing call) can gate which tier handles each conversation.

2. Conversation summarization

Instead of including the full transcript in every message, summarize the conversation after message 4 and pass a compressed summary (≈200 tokens) for messages 5+. This keeps context stable rather than growing linearly.

Effect: reduces average input per message from 1,350 to ~600 tokens — a 55% cut in input costs. On GPT-4o: $10,125 → ~$5,400/month.

3. Prompt caching

Your system prompt and FAQ context (400 tokens in our model) is repeated in every single message. Both Anthropic and OpenAI offer prompt caching — repeated prefix tokens are billed at 10–50% of the normal rate after the first request.

If your system prompt is 1,000 tokens and you cache it, you save 90% on those 1,000 tokens across all subsequent messages. On 100k chats × 8 messages = 800k messages, that's 800M tokens × $0.10/1M × 90% = $72/month saved on Flash alone. On GPT-4o the saving is larger in absolute terms.

What this means for your budget

Scale	Flash	Blended (routing)	GPT-4o
10k chats/mo (startup)	$43	$164	$1,013
100k chats/mo (growth)	$432	$1,639	$10,125
500k chats/mo (scale)	$2,160	$8,195	$50,625
1M chats/mo (enterprise)	$4,320	$16,390	$101,250

At 1M chats/month, the difference between Flash and GPT-4o is $97k/month — $1.16M per year. Even if Flash requires 15% more escalations to human agents, the math almost always favours the cheaper model tier.

Is quality good enough on Flash?

That's the real question, and it depends entirely on your use case. For FAQ answering, order status, return policy, shipping info — Flash performs comparably to GPT-4o in controlled evaluations. For nuanced complaint handling, multi-step troubleshooting, or situations requiring genuine reasoning about edge cases — frontier models have a measurable advantage.

The practical approach: run Flash for 30 days, measure escalation rate and CSAT, then compare to your baseline. If CSAT drops <1 point and escalation rises <5%, the model switch is justified by the cost saving.

Reference prices, July 2026. These assume standard public rates — enterprise pricing may differ. Verify on LLM price comparison or the provider's pricing page. Found an error? Report it →

FAQ

How much does an AI customer support chatbot cost per month?

At 100,000 chats/month with 8 messages average: Gemini Flash ~$432, GPT-4o mini ~$648, Claude Haiku ~$3,744, GPT-4o ~$10,125. The model choice is the biggest cost lever.

Why does conversation history make chatbots expensive?

Every message includes the full history as input tokens. Message 1 costs ~400 input tokens; message 8 costs ~2,500 — a 6× increase for the same output. Without conversation summarization, costs grow linearly with chat length.

Is Gemini Flash good enough for customer support?

For FAQ answering, order status, and policy questions — typically yes. For nuanced complaint handling or complex troubleshooting — benchmark first. Many teams run Flash for 60-70% of volume and escalate to a frontier model for edge cases.

How do I reduce AI chatbot costs?

Three main levers: (1) model routing — use Flash for simple queries, GPT-4o only for complex ones; (2) conversation summarization — compress old history into ~200 tokens; (3) prompt caching — both Anthropic and OpenAI offer 50-90% discounts on repeated prompt prefixes.