API Costs for AI Agents in a Contact Center: 2026 Guide

Disclosure: Dante AI publishes this guide. Dante AI plan figures come from our own pricing page and Help Center, read on 25 October 2026. We compare pricing models here, not vendors, and quote no other company's prices.

API costs for an AI agent in a contact center are more than the model's token bill. One AI reply can cost around a cent in tokens, but a working agent also pays for retrieval, hosting, voice minutes, phone lines and the engineering to keep it all running. Packaged platforms fold those costs into a seat, outcome, minute or credit price, so compare every option as a cost per resolved conversation.

What do API costs for AI agents actually cover?

When people ask what the API costs are for an AI-powered agent in a contact center or call center, they usually mean the language model. That is only the first of several layers you pay for when you build it yourself:

• Model tokens. Every reply is billed on the text the model reads (input tokens) and the text it writes (output tokens).
• Retrieval. Your help pages have to be split up, embedded and stored so the agent can look up the right passage before it answers.
• Voice. A phone agent also needs speech recognition, speech synthesis and a phone line, usually priced per minute.
• Hosting and monitoring. Servers, logs, alerts and someone on call when an answer goes wrong.
• Engineering time. Building and maintaining handover to a person, analytics, guardrails and retraining when your content changes.

A packaged AI agent platform bundles all five into one price. The question is which pricing model that price uses, and how it behaves at your volume.

How is contact center AI priced in 2026?

Contact center AI and contact center software are sold on five main pricing models. The same monthly figure can mean very different things depending on which one a quote uses.

Many quotes mix two of these, such as a platform fee plus a per-resolution charge. To compare like with like, convert every option to one number: what you would pay per resolved conversation at your real monthly volume. Our guide to contact center AI pricing walks through the same conversion for the four packaged models, and our containment rate guide explains how to count what an AI agent really resolves.

How much does one AI reply cost in API tokens?

Model providers quote token prices per million tokens, with input and output priced separately and output usually dearer. Rates differ a lot between light and premium models and change often, so always use your provider's current pricing page. For the arithmetic below we use round, illustrative rates of $2 per million input tokens and $10 per million output tokens.

A support reply reads far more than it writes, because the model is given your instructions, the help-page passages it retrieved and the conversation so far. Take an illustrative reply that reads 4,000 tokens and writes 300:

• Input: 4,000 tokens at $2 per million = $0.008
• Output: 300 tokens at $10 per million = $0.003
• Total: about $0.011 per reply

The rates, the 4,000 and the 300 are all assumptions for the arithmetic, not a quote or a measurement; your model, prompt, retrieval settings and answer length set the real numbers. Where a provider offers prompt caching, it can cut the input part when every reply starts with the same instructions.

Once you have a monthly figure, whether a plan price or an API bill plus everything around it, test it against what it saves with the calculator below.

What drives the cost of an AI agent up?

1. Model choice. The gap between a light model and a premium one is often ten times or more per reply. This is usually the biggest single lever.
2. Context length. Long instructions, many retrieved passages and long chat histories all add input tokens to every reply.
3. Turns per conversation. A conversation that takes six turns to resolve costs roughly six replies, and later turns read more history.
4. Voice minutes. Voice runs speech recognition, the model and speech synthesis continuously, so a minute of talk costs more than one text reply.
5. Phone lines. Phone numbers usually carry a monthly rental on top of usage.
6. Escalations. Every conversation handed to a person still costs that person's time, so a low resolution rate makes cheap AI expensive.
7. Maintenance. Content changes, new products and policy updates mean retraining and re-testing, whether you or the vendor does it.

What does the same workload cost on a credit plan? A worked example

Here is one invented business priced on Dante AI, a credit plan. Every plan figure below is from our pricing page and Help Center.

The business: an online retailer with 1,500 chat conversations a month, an average of 4 visitor messages each, plus about 150 minutes of AI voice calls.

1. Chat replies: 1,500 conversations × 4 replies = 6,000 replies. On the default model, GPT-5.4 Nano, one reply costs 1 credit, so that is 6,000 credits.
2. Voice: AI voice costs 5 credits per minute from the same balance, so 150 minutes = 750 credits.
3. Total: 6,750 credits a month.
4. Plan: Advanced is $120 a month with 8,000 credits, Human Handover to your team Inbox, and one rentable phone number for the voice agent. That leaves about 1,250 credits of headroom.
5. Cost per conversation: $120 / 1,500 chats is about 8 cents each, with the voice minutes inside the same price. The phone number rental is billed on top, and its monthly price is shown before you confirm it.

Now change one thing: the model. Standard models cost about 10 credits a reply, so the same 6,000 replies would need about 60,000 credits, more than Pro's 30,000 at $400. That is why model choice matters more than the plan name. The light default handles most help-page questions; keep the dearer models for the conversations that need them.

For comparison, the illustrative token arithmetic above puts 6,000 replies at about $66 a month in model tokens alone, before retrieval, voice, hosting, phone lines and engineering. Credit plans can cost more than raw tokens on paper and still cost less in total, because the other layers are included.

Credits refill every month and do not roll over. Extra credits are $20 a month per 1,000 on any paid plan. Yearly billing is ten times the monthly price.

How much does AI calling software cost?

AI calling software is usually priced per minute, sometimes with a monthly fee per phone number on top. To estimate yours, multiply your monthly call minutes by the per-minute rate, add the number rental, then divide by the calls the AI resolves without a person.

On Dante AI, voice conversations cost 5 credits per minute from the same balance as chat, so one bucket covers both channels. Free includes 10 minutes of voice a month. Advanced lets a voice agent answer 1 rented phone number and Pro up to 3. Phone numbers are currently available in the United Kingdom only, and calls are understood in English. See the AI Voice docs for the details.

Should you build on the API or buy a platform?

• Build on the API when you have engineers to own retrieval, voice, handover and monitoring, and a workflow no platform supports. You get the lowest unit price and the highest total effort.
• Buy a platform when you want the agent live this month and a bill you can predict. You pay a margin over raw tokens in exchange for everything around them.

Either way, test before you commit: list your 20 most common questions, run them through the agent, and work out your cost per resolved conversation. Our chatbot ROI guide turns that number into a payback estimate, and how much a chatbot costs to build covers the build-it-yourself path in more depth. To try the platform route, start free at dante-ai.com with Get started.

Sources

• Dante AI: pricing page, Understanding credits, Plans and pricing and AI Voice (read 25 October 2026).
• Claude API pricing page (read 25 October 2026), as one example of how a model provider publishes token prices: per million tokens, input and output priced separately with output dearer, and cheaper rates for cached prompts. We quote none of its prices.