A
Glossary
AI Agent Pricing
AI agent pricing is the set of pricing structures used to charge for autonomous AI agents, where the billable unit is a run, a task, a resolved outcome, or consumed tokens rather than a user seat. Agent workloads make cost vary per request, so the unit choice decides gross margin.
Key Takeaways
An agent's cost base has two dimensions, not one. Anthropic's Claude Managed Agents bills every session token at standard model rates and adds session runtime at $0.08 per session-hour, metered to the millisecond.
Per-seat pricing decouples revenue from cost. A single seat can trigger one agent run a month or ten thousand, and the invoice looks identical either way.
Outcome pricing shifts execution risk onto the vendor. If the agent retries five times before resolving a ticket, the customer pays for one resolution and you pay for five runs.
Most shipped agent products bill on a combination: a platform fee for access, metered runs or tokens for consumption, and a credit balance the customer draws down.
Caching changes the arithmetic enough to matter. In Anthropic's own worked example, a one-hour session drops from $0.705 to $0.525 once 40,000 of its 50,000 input tokens come from cache.
What units can you price an AI agent on?
The practical choices are per run, per task, per outcome, per token, and per unit of time, and each one moves risk between you and the customer.
Unit | What the customer pays for | Who carries execution risk |
Per run | Each agent invocation, successful or not | Customer |
Per task | A defined job, however many runs it takes | Vendor |
Per outcome | A confirmed result: ticket resolved, lead qualified | Vendor |
Per token | Model tokens consumed by the agent | Customer |
Per unit of time | Wall-clock runtime while the agent is working | Customer |
Per run and per token are the safest starting points because they track your cost almost exactly. Per outcome sells best and hurts most, since you absorb every retry. In my experience teams pick outcome pricing for the pitch, then quietly add a run cap or a credit balance behind it once the first month's margin report lands.
Why does agent pricing break token-based billing?
Token counts alone understate an agent's cost, because an agent consumes resources while it's thinking, waiting, and calling tools, not only while it's generating text.
Anthropic prices this explicitly. Claude Managed Agents charges on two dimensions at once:
All session tokens at standard per-model rates, with prompt caching multipliers applying normally
Session runtime at $0.08 per session-hour, accruing only while the session status is running and excluding idle time
Their published example makes the split visible. A one-hour Opus 5 session using 50,000 input and 15,000 output tokens costs $0.25 for input, $0.375 for output, and $0.08 for runtime, totalling $0.705. Price that session on tokens alone and you've left 11% of your cost off the invoice. Multiply that across a customer running agents continuously and the gap stops being rounding.
What does it cost to run agent pricing?
Running agent pricing costs you a metering layer that can attribute cost per run and per model, which is heavier than what subscription billing needs.
The pieces that have to exist before the pricing works:
Per-run cost attribution, so you can see which customer's agents are unprofitable rather than only the blended number
Model-level tracking, because the same run costs 5x more on a frontier model than on a small one
A stop mechanism, either a spending cap or a credit balance, since an agent in a retry loop can spend a month's budget in an afternoon
Reconciliation between what the model provider billed you and what you billed the customer
Skip the stop mechanism and the first runaway agent becomes a support credit. That's the failure I'd guard against first, ahead of any pricing sophistication.
Related reading
Further reading on pricing and metering agent workloads:
how to price ai agent usage based pricing walks through choosing a unit for an agent product.
how to pick the right pricing model for ai agent compares the model options side by side.
outcome based pricing covers charging for confirmed results instead of runs.
best ai cost tracking tools for agent infrastructure startups looks at tracking the cost side per agent.
best outcome based billing software for agent infrastructure startups reviews the tooling for outcome billing.
FAQ
Should AI agents be priced per task or per outcome?
Price per task unless you can control the retry rate. Task pricing still hides variable run counts from the customer, which is most of the appeal of outcome pricing, but it charges for the attempt rather than the result. Outcome pricing only works once you've measured how many runs a typical resolution takes and priced the average with margin on top.
Why does per-seat pricing fail for AI agents?
Seats don't correlate with agent consumption. One user can configure an agent that runs on every inbound email while another runs it twice a week, and per-seat pricing bills them the same. The revenue line stays flat while the cost line tracks agent volume, so margin erodes fastest on your most engaged accounts.
How do you price an agent that runs for hours?
Charge for runtime alongside tokens, the way Anthropic charges $0.08 per session-hour on top of token costs. Long-running agents hold compute and context open, and a token-only price misses that entirely. Bill the wall-clock time the agent spends working and exclude the time it spends idle waiting on input.
Do you need credits to run agent pricing?
No, but credits solve two problems at once: they cap exposure on a runaway agent and they let customers prepay for variable consumption they can't forecast. Metered postpaid billing with a hard spending cap achieves the same protection without a wallet, at the cost of a harder conversation when a customer hits the ceiling mid-month.
Back to glossary
