AI Answering Service Cost: The Model Is Nine Dollars

The short version: OpenAI’s voice model, GPT-Live-1, is priced at five cents a minute on OpenAI’s own developer pricing page, checked today. Run a small service business’s missed calls through it for a month, call it forty calls averaging four minutes, and the voice layer comes to eight dollars. Add the reasoning model doing the actual thinking and the whole thing lands under ten dollars. Bought as a finished product, an AI answering service costs $79 to $399 a month. That gap is not a markup on the model, and the line that will actually move your invoice is not minutes.

Start with that last point, because it is the one nobody prices out. Goodcall’s pricing page, checked September 14, 2026, says in as many words that it charges no fees for calls, minutes, or tokens. What it does meter is unique customers: 100 included on the $79 Starter plan, then fifty cents for each one after that. Work the arithmetic and there is an exact crossover at 200 unique customers a month, where Starter stops being the cheap plan and costs precisely what the next tier up costs. Goodcall’s page does not mention that number.

What OpenAI published, and what it leaves out

GPT-Live-1 reached the API on September 10 according to launch coverage. It is a full-duplex voice model, meaning it listens and speaks at the same time rather than waiting for you to stop talking, which is roughly the difference between a conversation and a phone tree.

The rate on OpenAI’s developer pricing page is five cents a minute, billed per second with no rounding up to whole minutes. A ninety second call genuinely costs seven and a half cents, not ten.

What that five cents does not cover is the thinking. The voice layer handles hearing and speaking. Whatever model decides what to say is billed separately at its own token rates, and so is every tool it calls to check your calendar or look up a customer. Most coverage of the launch stopped at the five cents. The bill does not.

A month of missed calls, priced out

Take a service business that sends about forty calls a month to voicemail. Not calls anyone is ignoring, just the ones that arrive at 6:40pm after the last van is back, or the second call that rings while you are already on the first, or the one that comes in while you are standing on a roof. Average them at four minutes.

The voice layer is the easy half. Forty calls at four minutes is 160 minutes, and 160 minutes at five cents is eight dollars. That figure is exact, because the rate is published.

The brain is an estimate, and here are the assumptions behind it, so you can argue with them. A four minute call runs roughly twelve exchanges. Each exchange resends the running transcript plus a system prompt holding your hours, services, prices and booking rules, call it 1,500 tokens. Cumulative input across one call therefore lands near 27,000 tokens, with about 1,000 tokens spoken back.

On gpt-5-mini, at $0.25 per million tokens in and $2.00 per million out, that is $0.0068 plus $0.0020, or just under nine tenths of a cent per call. Forty calls: 35 cents. On the full gpt-5, at $1.25 and $10.00, the same call costs 4.4 cents and forty calls come to $1.75. Prompt caching, which cuts repeated input by a factor of ten on both models, pushes it lower still.

So the month costs somewhere between $8.35 and $9.75. The striking part is not the total, it is the ratio. The voice layer is between 82 and 96 percent of the bill. The intelligence, the part everyone argues about online, is a rounding error.

What an AI answering service cost actually covers

Against that, the packaged prices look enormous. Goodcall’s Starter is $79 a month per agent, Growth is $129 and Scale is $249, with 15 percent off on annual billing. Slang.ai, aimed at restaurants, starts at $399 per location for Core and $599 for Premium, and publishes no minute allowances or overage terms at all.

Eight times to forty times the raw model cost sounds outrageous until you ask what the raw model cost actually buys, which is nothing you can answer a phone with. There is no phone number in it. Nothing writes to your calendar. Nothing texts a confirmation. Nothing decides what to do when the caller says something the agent was never told about, and nobody picks up when it books two jobs into the same Friday evening slot.

The model got cheap. The wiring around it did not, and the wiring is the product.

That reframes the comparison entirely. Per-minute rates and “unlimited minutes” badges are marketing about the cheapest input in the system. They tell you very little about what you are buying.

The line that moves your bill

Here is the arithmetic worth writing down. Goodcall’s Starter includes 100 unique customers and charges fifty cents for each additional one. Growth includes 250 for $129.

Set them equal: $79 plus fifty cents times (N minus 100) equals $129 when N is 200. At 200 unique customers a month, Starter and Growth cost exactly the same $129. Above 200, staying on Starter costs you money for nothing. Run the same calculation between Growth and Scale and the crossover is 490 unique customers, where both land on $249.

Which side of 200 you sit on has nothing to do with call volume and everything to do with what kind of business you run. A dental practice, a property manager, a regular-route service: mostly the same names every month, comfortably under the cap more or less forever. A locksmith, a mover, a tow operator, a restaurant taking reservations: almost entirely new names, over 200 quickly. And because the overage arrives in fifty cent increments rather than as a tier change, nothing announces it. The invoice just drifts upward.

What to compare instead

If the model is the cheapest part, stop shopping on it. The calls that decide whether this works are the ones the agent cannot finish.

So ask a vendor what happens on those. Does it warm transfer to a person, or only take a message? Can it say it does not know without inventing an answer or a price? Does it alert you immediately, or wait for a nightly digest? Can you read the transcript of every call it got wrong? Those answers vary enormously between products that look identical on a per-minute basis.

And be clear about what this is for. None of these calls are being answered by a person today. They go to voicemail, and most people calling a service business at 6:40pm do not leave one, they call the next number on the list. The job here is not to take the phone away from whoever answers it during the day. It is to stop the after-hours call from quietly becoming somebody else’s work order, and to hand your team the conversations actually worth a human. Measured that way, the right question about an AI phone agent is not how many calls it handles alone. It is how cleanly it gives up the ones it should.

For more on reading a 2026 pricing page, see our guide to what AI software is actually worth paying for, the $2 per conversation benchmark Salesforce set with Agentforce, and what changed when the cost of transcribing an hour of audio fell to ten cents.

Frequently Asked Questions

How much does an AI answering service cost for a small business?

Bought as a finished product, between $79 and $399 a month as of September 2026. Goodcall lists $79 for Starter, $129 for Growth and $249 for Scale, each per agent with 15 percent off annual billing, while Slang.ai starts at $399 per location. The underlying AI costs a small fraction of that: about eight dollars of voice time plus one to two dollars of reasoning for forty four-minute calls a month.

Why is the AI model so much cheaper than the answering service built on it?

Because the model is only the part that hears and talks. The price of a packaged service covers the phone number and call routing, the calendar and CRM integrations, the logic deciding when to escalate, call recordings and their retention, and support when something breaks at an inconvenient hour. None of that gets cheaper when token prices fall, which is why the gap between raw model cost and product price keeps widening rather than closing.

Does “unlimited minutes” mean there are no usage charges?

Not necessarily. Goodcall genuinely does not charge for calls, minutes or tokens, but it meters unique customers instead, including 100 on Starter and charging fifty cents for each one beyond that. For a business serving mostly the same customers month after month that cap is irrelevant, but for one talking to new people constantly it is the single biggest variable in the bill, and it arrives in small increments that are easy to miss.

Should a small business build its own AI phone agent instead of buying one?

Almost certainly not, unless you already have someone who maintains software for you. The five cents a minute is real, but you would be adding a phone number, telephony routing, calendar integration, escalation logic, transcript storage and an on-call human, which is precisely the work the $79 is paying for. Build it yourself only if you have unusual requirements no product covers, and price your own time honestly when you make that call.

Here is what we would genuinely like to know: if you have ever called a small business after hours and got a machine, did you leave a message, or did you just call the next one on the list?