What it costs to run support through an AI provider

Running support through an AI provider costs a fraction of a cent per reply, well under the penny our own landing page used to claim. Anthropic's own worked example lands at $37.00 per 10,000 tickets on their cheapest current model, before either discount below applies.

This page is one part of a larger guide. For the whole subject in one place, see running support with a coding agent.

Checked 22 August 2026, against Anthropic's, OpenAI's and Google's own published pricing pages.

What does a support reply actually cost in tokens?

A support reply costs tokens on both sides: tokens sent in, tokens sent back. What goes in is nearly all overhead, the instructions given to the agent, the knowledge base it draws on, and however much of the conversation came before, sent again unless something intervenes. The reply itself, the sentence a customer reads, is usually the smallest part. That's why the input price matters more than the output price for support, and why both discounts below attack the input side.

What do the major providers actually charge?

Eight models, three families, priced per million tokens.

The spread inside a family is worth noticing first. Haiku 4.5 prices input at $1 and output at $5. Sonnet 5 roughly doubles both, to $2 and $10, and Opus 5 doubles again, to $5 and $25. OpenAI shows the same shape: gpt-5-nano at $0.05 and $0.40 sits nowhere near gpt-5 at $1.25 and $10.

What does Anthropic's own worked example show?

Anthropic publishes a worked example on their pricing page, using Claude Haiku 4.5, their cheapest current model: total cost of around $37.00 per 10,000 tickets, at roughly 3,700 tokens a conversation. Divide the two and that's $0.0037 a reply, well under a penny, before either discount below is applied. That's the figure worth anchoring to, not the unattributed estimate this site's own landing page used to carry, never traced to a calculation. Correcting it is the reason this page exists.

If your own context runs longer than Anthropic's 3,700-token estimate, cost scales with it in roughly the same proportion: twice the tokens sent is roughly twice the cost. Worth doing that arithmetic on your own numbers rather than assuming Anthropic's estimate is yours.

Why does caching fit a support workflow so well?

A support workflow re-sends the same material on almost every request: the same system prompt, the same house style, often the same knowledge base article, fresh each time a ticket arrives. That's exactly what caching is built for. Anthropic's own description: a cache hit costs 10 per cent of the standard input price. OpenAI's cached input on the GPT-5 family is the same 10 per cent.

Run the arithmetic on Haiku 4.5's $1 input price and a cache hit costs $0.10, a tenth of standard, for everything sent before that hasn't changed. Only genuinely new material gets billed at the full rate, which is why support is close to the best possible fit for this discount.

What does the Batch API change?

Batch processing trades speed for price: submit work that doesn't need an answer in the next few seconds, and it comes back within a window instead of instantly, at half the standard cost on both input and output tokens. Anthropic's own description is "a 50% discount on both input and output tokens." OpenAI and Google publish the same 50 per cent figure.

Applied to the $37.00 anchor above, work that can wait, an end-of-day pass over routine tickets rather than a live chat, roughly halves that figure again. Anything a customer is watching in real time still needs the standard price, but a queue of lower-priority tickets processed overnight is exactly the shape batch pricing was built for.

What makes the real bill bigger than the arithmetic suggests?

Four things, honestly.

A growing thread re-sends its own history. Without caching, a ticket that goes back and forth five times means the model re-reads the first four exchanges every time, so cost per reply climbs as the thread gets longer.

Retries are billed like any other request: a call that fails and runs again costs the same as the one that worked, now paid for twice.

A bigger model than the job needs is the largest lever of all. The table above shows Opus 5 at roughly five times Haiku 4.5's price, on both input and output. Defaulting to the strongest model for every reply, rather than the cheapest one that can do the job, is usually the biggest cost decision in the system, made once in configuration.

And the cost that never appears in a token count: a person reading the draft, deciding whether to send it, rewriting it when it's wrong. Where a human reviews every reply, that's often the larger cost of the two.

How does this compare to a subscription?

Token cost and subscription cost behave differently, and that's the real comparison, not the headline numbers.

A per-tracked-user subscription like Canny.io's rises in steps and never falls: entry is $79 a month billed yearly, or $99 monthly, for 100 tracked users, and everyone who posts, votes, comments counts toward that number permanently. At 1,000 tracked users the same plan is $529 a month billed yearly, or $661 monthly. A per-seat plan like Featurebase's, $29 a seat a month billed yearly, behaves the same way in a different dimension: add someone to the team and the bill goes up, whether or not they answered a single ticket.

Token cost has no floor like that. It scales with tickets actually answered, at a fraction of a cent each, and doesn't accumulate a permanent headcount. It has no ceiling either: enough volume, a big enough model, no caching or batching, and the bill keeps climbing with no upper tier to cap it. The honest comparison isn't which is cheaper in the abstract, it's which shape fits how the work happens. What feedback tools cost covers entry pricing across the wider category.

What about Gemini's free tier?

Gemini is the one provider here with a genuine free tier, not only a trial, which matters if the budget is close to zero rather than merely small. The trade is on Google's own page: free-tier data is used to improve their products, paid-tier data is not. Worth knowing before choosing free to save a fraction of a cent paid would have cost anyway.

The AI provider is usually the only line in a self-hosted setup that scales with use rather than staying fixed. What a self-hosted support centre costs covers the rest of the bill.

Frequently asked questions