What an AI support answer actually costs in 2026

A single AI support answer, priced at public list rates in 2026, costs somewhere between well under ₹1 and a few ₹ — depending almost entirely on which model you use and whether the answer needs tools. That is the whole headline, and you can verify every number yourself from published price pages. This post shows the arithmetic, because once you have seen it you will never again read "$0.99 per resolution" the same way.

Everything below is reader-computable public math. We are not going to tell you what anything costs us to run — that is our business, and yours is the formula. An AI answer is not a mysterious cloud expense. It is a token count times a price per token, and both numbers are published.

The anatomy of one answer

Front-load the shape: a support answer is mostly reading, not writing. The model consumes a large input and emits a small output, and that asymmetry drives the cost.

A realistic support answer feeds the model four things. A system prompt — the instructions defining how the assistant behaves — is commonly 1,000 to 2,000 tokens. The knowledge context — the help articles retrieved to ground the answer — is the big one, often 2,000 to 4,000 tokens. The conversation history adds a few hundred to a thousand. And the model writes a reply of maybe 150 to 400 tokens. Round it to a clean example: about 5,000 tokens in, about 300 tokens out. A support answer reads a small booklet and writes a sticky note — you pay mostly for the reading.

The same answer, three models

Now price that 5,000-in / 300-out answer at three public list tiers, as of mid-2026, at roughly ₹95 to the US dollar.

A frontier model — GPT-4.1, listed at $2.00 per million input tokens and $8.00 per million output. The input costs 5,000 × $2 ÷ 1,000,000 = $0.010. The output costs 300 × $8 ÷ 1,000,000 = $0.0024. Total: about $0.0124, or roughly ₹1.2 per answer.

A value model — GLM-5.2, listed at $1.40 input and $4.40 output per million. Input $0.007, output $0.00132, total about $0.0083 — roughly ₹0.80 per answer.

A small model — GPT-4.1 mini class, around $0.40 input and $1.60 output per million. Total about $0.0025 — roughly ₹0.24 per answer, well under a rupee.

There is the range in one table's worth of arithmetic: the same support answer costs about ₹0.24, ₹0.80, or ₹1.20 depending purely on the model tier. A frontier answer and a small-model answer are the same question — they differ by a factor of five in price, not a factor of a hundred. For a plain, well-grounded FAQ answer, the cheap model is often indistinguishable to the customer; the frontier premium buys headroom for the hard, ambiguous cases.

Tools change the number by multiples

Everything above assumes one model call. The moment the answer needs to do something — look up an order, check a subscription, take an action — the math multiplies, because each tool step is another round trip that re-sends the growing context.

A tool-using answer might call the model three or four times: read the question, decide to call a tool, read the tool result, answer. Each call re-sends the system prompt and accumulated context, so the total input balloons from 5,000 tokens to 15,000 or 25,000. At frontier rates that pushes a single resolved conversation from about ₹1.2 toward ₹4 to ₹5. A question the AI can answer from an article costs a rupee. A question where it has to go look something up costs several — because it reads the booklet three times, not once. This is why "resolution rate" and "cost per resolution" are not the same curve: the answers that need tools are both the most valuable and the most expensive.

Why caching quietly rewrites the bill

Here is the lever that separates a naive deployment from an efficient one: most of that 5,000-token input is the same on every answer. The system prompt does not change. The most-used help articles do not change hour to hour. Providers price cached input far below fresh input — GPT-4.1's cached rate is listed around $0.50 per million versus $2.00, and GLM-5.2's around $0.26 versus $1.40. That is a 75-to-80 percent discount on the repeated portion.

For a support workload — where the same policy paragraphs and the same system prompt get read on nearly every conversation — that is not a rounding error. The cheapest token is the one you already paid to read; caching is how a support system stops paying full price for the same paragraph a thousand times a day. A well-built system caches aggressively, which is exactly why raw per-answer cost in production trends toward the low end of the ranges above, not the high end.

So why does retail cost 99 cents?

Put the two numbers side by side. A grounded support answer costs on the order of ₹1 in raw model tokens. A per-resolution retail price like $0.99 is, at ₹95 to the dollar, about ₹90 to ₹95 — roughly a hundred times the raw token cost. That gap is not a scandal; it is the product.

The difference pays for everything the raw model does not do on its own: the retrieval and grounding that finds the right article, the safety and guardrails that stop it inventing a refund policy, the evaluation harness that certifies every model change before a customer sees it, the human-in-the-loop for money-moving actions, the infrastructure, the uptime, the support behind the support. You are not paying a hundred times the token cost for tokens. You are paying for the machine that makes the tokens trustworthy. A raw model call is cheap and confidently wrong; a support product is the apparatus that makes it cheap and reliably right.

What this means at a month's scale

Zoom out from one answer to a month, because that is the number you actually budget. Take a small business fielding 3,000 conversations a month where the AI handles, say, 60 percent — 1,800 answers. At the value-model rate of roughly ₹0.80 an answer, that is about ₹1,440 in raw model tokens for the entire month. Even if a third of those answers use tools and cost several rupees each, you are still looking at raw model spend in the low thousands of rupees a month for a business doing thousands of conversations. A month of grounded AI answers, priced at the raw tokens alone, costs less than a single day of a human agent's time. That is the fact that makes support automation economically obvious — and also the fact that explains why the retail gap exists, since the tokens were never the expensive part of making the answer trustworthy.

Compute your own number

The formula is short. Take your average answer's input tokens (system + knowledge + history) and output tokens, multiply each by your chosen model's published input and output price, and add them. Multiply the input side by roughly 0.2 to 0.3 for the cached portion. Multiply by three or four for answers that call tools. That is your real per-answer cost, and it will almost always land under a few rupees.

That token-anatomy view is exactly how we think about Iris under the hood: keep the grounding tight so the knowledge context stays lean, cache the parts that repeat, and reserve the expensive frontier calls for the answers that genuinely need them — so the AI can answer from your own help content at a cost measured in small change, and the price you pay reflects the harness around the model, not the raw tokens alone.