What an AI conversation actually costs

Contents

Every AI vendor quotes you a price per conversation. Almost none of them will show you how they got there, which is usually because the number is a guess with a margin on top.

We compute ours per call, in cents, and store it against the record. Here's the arithmetic and, more usefully, where the money actually goes.

Start with the published rates

Token pricing is public. Here are the rates our cost table carries, in US dollars per million tokens:

Two things worth noticing straight away.

Output costs four to five times what input costs, across every provider. That ratio drives more design decisions than the headline price does. A prompt that returns structured JSON with six fields is dramatically cheaper than one that returns a paragraph of prose, and in most of our pipelines the paragraph was never needed.

And the spread between the top and bottom of that table is roughly a hundred to one. A pipeline that routes everything to the largest model because it's easiest costs two orders of magnitude more than one that classifies with a small model and reserves the large one for the step that needs judgement.

We compute it ourselves, because nobody hands it to you

A detail that surprises people building this for the first time: the SDKs don't tell you what a call cost. The Anthropic SDK doesn't report cost. OpenAI's usage object carries token counts with no pricing attached. You get the tokens and you're on your own.

So we keep a rate table in the codebase, look up the model, and do the multiplication locally. Cost lands in whole cents on the pipeline record alongside the prompt and completion token counts.

Two details in there we'd have got wrong without production traffic. Model names arrive with dated suffixes, so a lookup falls back to a longest-prefix match across known families. And an unknown model returns zero rather than throwing, because a pricing gap should show up as a warning in the log and a nought in the report, never as a failed customer conversation.

The table only carries models we actually serve. Every row is verified against the provider's published rates when it's added, and the update pattern is: a new model gets a row before it gets traffic.

Now the part the token table doesn't show

If you're costing a text pipeline, tokens are most of it. If you're costing a voice call, tokens are a rounding error, and this is where most published price-per-conversation figures quietly fall apart.

A three-minute voice conversation on Switchboard spends money in five places:

Telephony. Per-minute inbound or outbound, and it runs for the whole call, including the silence while the customer finds their policy number.

Speech to text. Streaming transcription, billed by audio duration. Also running for the entire call, including that same silence.

Text to speech. Billed by characters synthesised. This is the one you can genuinely influence, because it's driven by how much your agent talks. An agent that says "Let me just look into that for you, this will only take a moment" on every single tool call is charging you for its own filler.

The language model. Billed per turn, and here's the trap: cost per turn grows through the call. Turn twelve resends the whole conversation so far as input. A twenty-turn call doesn't cost twenty times turn one, it costs considerably more, because the input side compounds.

Everything the caller didn't hear. Retries, failover between providers, the classification and routing calls that happen before anyone speaks. Real spend that never appears in a naive estimate.

Total call duration multiplies three of those five lines. Which means the single most effective cost control in voice AI has nothing to do with model choice. It's not making the caller wait.

Where this changes what you build

Once cost is measured per call rather than estimated per month, some decisions stop being arguments.

Model routing gets easy to justify. Classification, extraction and routing go to a small model. The step requiring actual judgement goes to a large one. Everybody agrees with this in principle and nobody does it without numbers, because the small model is slightly worse at everything and there's no visible reward for the trade until you can see the bill by pipeline step.

Structured output pays for itself twice. Cheaper on the output side, and easier to validate, which means fewer retries. Retries are pure loss: you pay for the failed call and the successful one.

Context management becomes a cost lever rather than a quality one. Summarising a long conversation instead of resending it verbatim is normally argued on quality grounds. The cost argument is simpler and lands better with a finance director.

Long calls look different. A caller stuck in a loop is costing you telephony, transcription and a compounding model bill simultaneously. Escalating to a human at eight minutes can be cheaper than the AI eventually getting there, before you count what the caller thinks of you.

What to ask a vendor

If you're comparing AI platforms, four questions separate the ones who've measured from the ones who've guessed.

Can you show me cost per conversation for last month, broken down? If the answer is a flat per-minute or per-conversation rate, they're carrying the variance and pricing in a margin for it. That's a legitimate business model. It's not transparency, and you'll pay for the calls that went well.

Which model handles which step? "We use GPT-4o" for the whole pipeline means nobody has looked at the bill.

What happens to cost as the conversation gets longer? If they haven't thought about input compounding, they haven't run a long call.

Do you record token counts per turn? You can't optimise what nobody stored. We keep prompt and completion tokens on the pipeline record precisely so the question stays answerable six months later.

The short version

AI conversation pricing is knowable. The published token rates are the easy part and, for voice, the smallest part. Duration drives the bill, output tokens cost several times input tokens, and the invisible calls are real.

Measure it per call, store it with the record, and the decisions about which model goes where mostly make themselves.

---

Further reading:

• AI resilience: why we route across multiple providers, and what that does to cost
• The email the AI answers on its own: where the confidence tiering sits in the same pipeline
• Pricing: how SwiftCase itself is priced
• Switchboard: the AI agents layer this all runs inside