AI support agents bill a median of $0.89 per resolved conversation. The model behind them costs cents

Almost every support tool now sells an AI agent, and most of them bill it by the result: a fixed fee each time the AI resolves a conversation. On 2026-10-10 we read the pricing page of 20 of these products and the price list of the four main model providers. This article puts the two side by side: what you are charged per resolution, and what the underlying model costs for the same conversation.

What vendors charge per resolution

9 of the 20 products publish a price for one resolved conversation, or for the nearest thing they bill. They sit in a narrow band, from $0.49 to $1.50, with a median of $0.89.

The unit is where the money is

• Outcome, not resolution. Intercom bills an "outcome", which is a resolution but also a handoff to a human that follows a procedure you configured, or a disqualification. At most one per conversation.
• Silence counts as success. Several definitions treat a conversation as resolved when the customer does not come back within a window: 72 hours at HubSpot, 24 to 72 hours at Front.
• Sessions, whatever happens. Freshworks bills each session with the AI agent, resolved or not.
• Double counting. At Gorgias an automated interaction is billed and also uses one of the helpdesk tickets in your plan.
• Levels. Front charges a conversation once at the highest level reached: 5 cents if the AI only triaged it, 39 cents for a handoff, 89 cents for a resolution.

The other pricing models

Per-message pricing is far cheaper per unit and puts the risk on you: a long conversation costs more, a resolved one costs the same as a failed one. Per-resolution pricing does the opposite. Neither is the cost of running the model.

What the model costs

A language model is billed by the token: so much per million tokens read, so much per million written. These are the standard list prices of the current models on 2026-10-10.

To turn that into a cost per conversation you need token counts. We have one measured figure and two assumptions. The measured one comes from our own product: across 171 reply drafts generated in muro between May and September 2026, the median call read 556 tokens and wrote 154. Our knowledge bases are small, so we added two heavier cases and state them as assumptions.

Side by side

On a mid-range model, the long conversation costs 15 cents in tokens; the median vendor price for resolving it is $0.89, about 6 times more. On a small model the same conversation costs 0.74 cents, about 120 times less than the median fee. Only the most expensive flagship model, used for every reply of a very long conversation, gets within range of the fees, at 74 cents.

What to do with this

1. Price your own volume. Multiply your monthly conversations by a realistic resolution rate and by the fee, then compare with a flat plan or with paying the model directly.
2. Ask how a resolution is counted before you sign: abandonment window, handoffs, one or several per conversation.
3. Check what happens to failures. Per-resolution pricing means the vendor carries unresolved conversations; per-message pricing means you do.
4. Look for a bring-your-own-key option. If the product lets you plug in your own model account, you pay the list prices above and the vendor charges for the software only.
5. Re-check the model prices every few months. On the same price page, Claude Haiku 4.5 is still listed at $1 and $5 per million tokens; its successor Claude Haiku 5.5 is at $0.1 and $0.5.

For transparency about our own product: muro includes a monthly pool of AI credits in each plan, one credit per AI action, and lets you connect your own Anthropic key, in which case AI usage is unlimited and billed by Anthropic at its list price with no markup from us. That is the model this article makes look good, so weigh it accordingly.

Method and limits

• Vendor prices. Read on each vendor's own pricing or help page on 2026-10-10, from France. Where a page shows several rates (committed and pay-as-you-go, annual and monthly), the chart uses the lower published one and the table gives both.
• Zendesk. Its pricing page still lists $1.50 and $2.00 per automated resolution, while its help centre says a tiered model introduced in May 2026 replaces them without publishing the tiers.
• Model prices. Standard tier, short context, from the official price pages. Caching, batch processing and volume agreements all reduce them.
• Token counts. One measured value from our own production (a reply draft, median of 171 calls) and two stated assumptions. Your conversations may be longer; the dataset lets you redo the arithmetic with your own figures.
• Not measured. Resolution rates, answer quality and the engineering needed to run a model yourself. Those decide whether any of this is a good deal.

Sources

1. 1Freshworks Freddy AI Agent, pricing (accessed 2026-10-10)
2. 2HubSpot customer agent, pricing (accessed 2026-10-10)
3. 3Help Scout AI Answers, pricing (accessed 2026-10-10)
4. 4Re:amaze AI Agent, pricing (accessed 2026-10-10)
5. 5Front Autopilot, pricing (accessed 2026-10-10)
6. 6Gorgias AI Agent, pricing (accessed 2026-10-10)
7. 7Intercom Fin, pricing (accessed 2026-10-10)
8. 8Text (LiveChat) AI Agent, pricing (accessed 2026-10-10)
9. 9Zendesk AI agents, pricing (accessed 2026-10-10)
10. 10Chatwoot Captain, pricing (accessed 2026-10-10)
11. 11tawk.to AI Assist, pricing (accessed 2026-10-10)
12. 12Chatbase, pricing (accessed 2026-10-10)
13. 13SiteGPT, pricing (accessed 2026-10-10)
14. 14Anthropic, API price list (accessed 2026-10-10)
15. 15OpenAI, API price list (accessed 2026-10-10)
16. 16Google, API price list (accessed 2026-10-10)
17. 17Mistral AI, API price list (accessed 2026-10-10)