AI Agent Costs: What Makes the Monthly Bill Go Up
Raise your agent’s escalation rate from 5% to 15% and its modelled monthly bill roughly doubles. The token line doesn’t move.
IN 30 SECONDS
An agent’s escalation rate is the share of its runs it hands to a person for review. In this post’s model, that rate moves the monthly bill far more than the token price, and cost per agent can climb once agents share parts.
KEY TAKEAWAYS · 3
- In the modelled example, raising escalations from 5% to 15% of runs takes the bill from $521 to $1,121 a month. Tokens stay at $26.
- Cost per agent can climb once agents share parts, such as one model or one search index, because a single change upstream can then touch many agents.
- There is no universal agent count at which costs change shape. Measure migration and review effort on your own fleet.
On this page
Escalations move the bill, not tokens
IN SHORT Same modelled agent, three times the escalations: the bill roughly doubles and the tokens stay at $26.
WORDS TO KNOW
- Escalation
- The agent handing a case it is not sure about to a person, with what it found.
- Modelled cost
- A cost worked out from assumptions written down in advance, not read off a real invoice.
- Migration
- Moving agents onto a new version of a shared part, such as a new model, and testing them again.
Take the read-only assistant from stage 1, with the same flat platform, data and engineering lines. Each Escalation The agent handing a case it is not sure about to a person, with what it found. See the glossary costs 3 minutes of review at $60 an hour. Every figure is a Modelled cost A cost worked out from assumptions written down in advance, not read off a real invoice. See the glossary, not an invoice, and only the escalation rate changes below; routine user reading stays out, and your shares will differ.
A WORKED EXAMPLE
Illustrative numbers
PASSBase case: 5% of runs escalate
INPUT
- Runs per month
- 2,000
- Escalation rate
- 5%
- Review
- 3 min at $60/h
WHAT HAPPENS
- 5% of 2,000 runs gives 100 escalations.
- 100 reviews of 3 minutes make 5 hours, so $300.
- Tokens stay at $26. The other three lines add $195.
OUTPUT
- Review
- $300
- Total
- $521 a month
ABOUT 2× THE BILLSame agent, 15% of runs escalate
INPUT
- Runs per month
- 2,000
- Escalation rate
- 15%
WHAT HAPPENS
- 15% of 2,000 runs gives 300 escalations.
- 300 reviews of 3 minutes make 15 hours, so $900.
- Tokens are still $26. Nothing about the model changed.
OUTPUT
- Review
- $900
- Total
- $1,121 a month
IN PLAIN WORDS
Tokens never moved, and the total roughly doubled. What moves the bill is the escalation rate, a design choice, not a vendor choice. Haggle over the model and you’re fighting over a twentieth of the base bill.
Set an agent’s bar for “I’m not sure” too cautiously and you’ve quietly hired someone.
Where that bar sits, and how it’s scored before a release, is the job of the eval suite, the set of tests that score an agent.
Coming up next: And why would agent 40 cost more to carry than agent 4?
Why agent 40 costs more than agent 4
IN SHORT Cost per agent stays nearly flat for a while, then can climb once agents start sharing the same parts.
Volume doesn’t cause the climb. A Shared dependency A part many agents rely on at once, such as one model, one search index or one login path. See the glossary does: one model, one search index, one prompt template or one login path that many agents use. Change it once and you may need to retest every agent that uses it.
The stages below are illustrative, not measured breakpoints from the Integra fleet. Your curve depends on what’s shared, how often it changes and who owns recovery.
STEP BY STEP
Step 1. Agents 1 to 3. One engineer may hold most of the context, so upkeep is nearly invisible. Write down owners and recovery steps early.
- Agents 1 to 3 One engineer may hold most of the context, so upkeep is nearly invisible. Write down owners and recovery steps early.
- Around agent 6 Nobody can say from memory what each agent does or who owns it. The first fleet tool appears: a list.
- Around agent 15 An index, a prompt template, a model and a login path are now shared. Each new agent is cheap to build and costly to carry.
- Around agent 30 Migrations, shared tests and the list no longer fit between features. Someone now owns the fleet, named or not.
- At about 46 Count owners, not agents: how many people can handle recovery and keep the shared parts working?
WATCH OUT
There is no universal agent count at which costs change shape. Measure Migration Moving agents onto a new version of a shared part, such as a new model, and testing them again. See the glossary and review effort on your own fleet.
60 days
minimum notice before a public model is retired
Source: Anthropic, model deprecations page.
Model changes aren’t hypothetical. Anthropic gives customers with active deployments at least 60 days’ notice before it retires a publicly released model. The work that follows depends on compatibility, test coverage and rollout; fleet size alone can’t tell you how long a migration takes.
ITEM GET
Shared parts multiply upkeep
FAQ
Every agent needs an owner. A larger or more connected fleet may need someone who owns the platform. Decide from the maintenance work and the operational risk, not from an arbitrary agent count.
STAGE CLEAR
What you now know
- Why, in this model, the escalation rate moved the bill far more than the token price.
- Why cost per agent can climb once agents share a model, an index or a login path.
Check yourself
Tap or click a question to see the answer.
Illustrative case: in fleet A, one agent out of 20 uses model M; in fleet B, all 20 do. The provider retires M. Which fleet has more agents to check?
Fleet B: 20 agents depend on M, against one in A, and each may need retesting. How long that takes still depends on compatibility, tests and rollout.
Former CTO of Integra. Named author (1 of 5) of ERC-3643, first author of ERC-6960, co-author of ERC-7410, and co-author of ERC-8203, which is still a draft. Building production AI and regulated Web3 systems.
Have a project to ship or a role to fill?
Discuss your project or roleFollow the engineering notes
Get new articles on production AI and smart contracts by email.