What a RAG chatbot actually costs, with the numbers you can check

Search "how much does a RAG chatbot cost" and you'll get pages of ranges. Some start in the low thousands, some end in the hundreds of thousands, and almost none say where the numbers came from. Most are written by agencies whose own quote is somewhere in the middle.

We build RAG systems for clients, and we run one of our own in production: ConGreeto, a website assistant that answers from each customer's own site and catalogue. So this post doesn't do ranges. It covers what each layer of a RAG chatbot costs, which numbers you can check yourself, and how to choose between buying a product, building it yourself, or having it built.

Short version: the model bill is usually the smallest line in the budget. The expensive part is the work that makes answers correct, and keeps them correct.

The three cost layers

Every RAG chatbot, however it's delivered, has three kinds of cost:

1. Build. Ingestion (getting your content in), retrieval, the prompt, the chat interface, lead capture or handoff, and the admin tools around it. This is people's time.
2. Run. Model tokens, embeddings, a database with vector search, hosting and background workers. This is the part most guides obsess over.
3. Maintain. Re-ingesting content when it changes, fixing wrong answers, testing after model updates, handling the edge cases real users find. This never ends, and it's the line most budgets leave out.

Products, DIY and agencies split these three layers differently. That's the whole decision.

Running cost, measured

Here's the run layer with numbers you can verify.

Model tokens

OpenAI's published API prices (standard tier, checked 11 October 2026):

Model

Input, per 1M tokens

Cached input

Output

gpt-4o-mini

$0.15

$0.075

$0.60

gpt-5.4-mini

$0.75

$0.075

$4.50

text-embedding-3-small

$0.02

—

—

Our measurement: across 171 recorded replies from ConGreeto, the total tokens per reply (prompt plus answer, from the API's own usage figures) averaged 3,757, with a median of 3,048 and a maximum of 6,873. That was test traffic before launch, not customer conversations, and the total doesn't separate input from output.

So instead of guessing the split, take the worst case: price every token at the output rate, which is the most expensive one.

Per reply, worst case

Per 1,000 replies, worst case

gpt-4o-mini, average reply (3,757 tokens)

$0.0023

$2.25

gpt-4o-mini, our largest reply (6,873 tokens)

$0.0041

$4.12

gpt-5.4-mini, average reply

$0.017

$16.91

Real costs are lower than this table, because most of a RAG prompt is input, and a well-ordered prompt gets a large share of it billed at the cached rate. We wrote up how prompt block order changed our cached share separately.

The point stands either way: on a small model, a thousand replies cost a few dollars at most. Model choice matters much more than prompt trimming. The same reply on gpt-5.4-mini costs several times more.

Embeddings

Embedding is cheaper still. One customer website we crawled (291 pages) produced about 7,200 chunks. Our chunks average about 104 tokens, so that's roughly 750,000 tokens. At $0.02 per million, embedding the whole site cost about 1.5 cents.

Re-embedding is where waste creeps in, not the first crawl. If every re-crawl re-embeds every page, the cost multiplies by every page and every crawl. Hashing content and caching embeddings by text fixes that (also covered in the crawl post).

What the per-reply number leaves out

Be wary of any cost estimate that stops at "tokens per reply". A production chatbot makes other model calls besides the answer itself:

• embedding each visitor question to search the knowledge base,
• summarising long conversations so they fit in the prompt,
• scoring or classifying the conversation (is this a lead, how warm),
• extracting structured facts (prices, services) from crawled pages.

Each one is small. Together they can rival the answer itself, and most of them aren't visible on a per-reply dashboard. Ask any vendor or agency which of these their estimate includes. Our AI agent running-cost post goes deeper on this.

Infrastructure

A database with vector search (we use PostgreSQL with pgvector), an API server, background workers for crawling and embedding, a queue, and somewhere to run a headless browser for JavaScript sites. At small scale this is a modest monthly hosting bill. It grows with crawl volume and memory, not with chat volume. Our multi-tenant architecture post covers where the real limits show up.

Option A: buy a product

Prices from each vendor's own pricing page, checked 11 October 2026:

Product

Pricing model

Entry point

Chatbase

Monthly plan with message credits

Hobby $40/month (700 message credits); Standard $150/month (4,000 credits); 20% off yearly

Intercom Fin

Per outcome

$0.99 per outcome, 50 outcomes per month minimum, works with other helpdesks

Intercom (helpdesk + Fin)

Per seat + per outcome

From $19 per seat/month (billed annually) plus $0.99 per Fin outcome

ConGreeto (our product, so weigh accordingly)

Monthly plan with conversations

From $37.50/month billed annually; built for lead capture and qualification rather than support

What you get: build and run costs bundled into one predictable bill, a working product this week, and maintenance done by someone else.

What you give up: control over retrieval, prompts and data handling, and you're limited to the integrations on the vendor's list. Per-outcome pricing is easy to start with but scales directly with success. At $0.99 an outcome, 2,000 outcomes is $1,980 a month.

Choose this if: your use case is common (support, FAQ, lead capture on a website) and a product already does it well.

Option B: build it yourself

The tutorial version of RAG takes an afternoon. The production version takes much longer, and the time doesn't go where people expect. In our own build, the expensive problems were:

• Ingestion that silently fails. Our first crawler indexed only the home page of JavaScript-rendered sites and reported success.
• Content that changes. Re-crawls, change detection, and deciding what wins when a human's correction and the website disagree.
• Memory and queues. One large site took down a worker before we batched embedding.
• Prompt behaviour. Answering contact questions directly, refusing what the bot shouldn't answer, handing off cleanly.
• Measurement. Knowing tokens per reply, cache hit rate and answer quality before optimising anything.

None of that is model cost. All of it is engineering time, and most of it shows up after launch. We documented the five failures in detail.

Choose this if: the chatbot is core to your product, you have engineers who will own it for years, and no product fits your data, integrations or compliance needs.

Option C: have it built

An agency or contractor takes the build layer and often part of maintenance. Quotes vary enormously because the scope does. A quote is driven by:

• Sources. One website is a different job from a website, a document library, a product database and a CRM.
• Accuracy bar. "Usually right" is cheap. "Must cite its source and refuse when unsure" needs evaluation sets and testing.
• Integrations. Each system it reads from or writes to (CRM, booking, helpdesk) is real work.
• Languages and channels. Multilingual, right-to-left scripts, WhatsApp, voice.
• Who runs it afterwards. A handover to your team, or ongoing operation by the agency.

Red flags in a quote:

• No mention of how content is kept up to date after launch.
• No plan for measuring answer quality, only for "building the bot".
• Running costs stated as tokens only, with no infrastructure, re-indexing or extra model calls.
• A fixed price before anyone has looked at your content sources.
• You don't own the data, the prompts or the code at the end.

Choose this if: your needs are specific enough that no product fits, but you don't want to hire and keep a team for it.

Which one fits

Your situation

Best fit

Standard use case, small team, want it live this week

Buy

Lead capture and qualification on a business website

Buy a product built for that

Unusual data sources, integrations or compliance rules

Have it built

Chatbot is the product, long-term in-house team

Build

Not sure yet

Buy first. Real usage tells you what to build later

Running a product for a month is the cheapest way to learn what you actually need. Teams that start by asking for a custom build often discover a product covers most of it. The ones that still need custom work then write a much better brief.

Frequently asked questions

How much does a RAG chatbot cost to run per month?

On a small model like gpt-4o-mini, the model bill per reply is a fraction of a cent. Our worst-case figure is $0.0023 per average reply, so even tens of thousands of replies a month cost tens of dollars in model tokens. Hosting, background workers, and the extra model calls around each reply usually cost more than the replies themselves.

Is it cheaper to build a RAG chatbot or buy one?

For common use cases, buying is almost always cheaper once you count engineering time and maintenance. Building pays off when the chatbot is central to your product or no product fits your data and integrations.

Why do agency quotes for RAG chatbots vary so much?

Because the scope does. Number of content sources, accuracy requirements, integrations, languages, and who runs it afterwards each change the work involved. A quote without a look at your content sources is a guess.

Do embeddings cost a lot?

No. Embedding a 7,200-chunk website cost us about 1.5 cents. The cost risk is re-embedding unchanged content on every re-crawl, which change detection and caching prevent.

Sources

• API pricing: OpenAI, standard tier, checked 11 October 2026.
• Chatbase pricing, checked 11 October 2026.
• Fin pricing and Intercom pricing, checked 11 October 2026.
• ConGreeto pricing, checked 11 October 2026. ConGreeto is our own product.
• Tokens per reply: 171 replies recorded from ConGreeto's API usage figures, August 2026, pre-launch test traffic. Chunk counts and sizes: ConGreeto production crawl data.

From us

• Building a Chatbot Knowledge Base From a Website: What the Crawl Gets Wrong
• What an AI Agent Actually Costs to Run in Production
• Multi-Tenant AI SaaS Architecture
• RAG implementation services: if your needs fall under "have it built", tell us about your content sources and we'll scope it honestly, including when a product would serve you better.