DEV Community

weiwuji
weiwuji

Posted on

The Enterprise Agent OS Price Sheet: Seats, Actions, and Compute Are Fighting

The Pain

The budget is approved, the Agent is live, and the month-end bill makes no sense. You bought 50 seats. Why is there also a credits line, a token line, and a runtime line?

What You'll Learn

  • The three billing units used by enterprise Agent OS products.
  • The exact public prices behind seats, actions, and compute.
  • Why the visible price is only one slice of the real cost.
  • Three questions to answer before you choose a pricing model.

Three pricing units

Seats, actions, and compute are three different businesses wearing the same "Agent" label.

What "Agent OS" means here

An enterprise Agent OS is not a single product. It is the layer that manages Agents: Microsoft Agent 365, Google Gemini Enterprise, Salesforce Agentforce, and AWS Bedrock AgentCore all sell a version of that layer.

The important difference is not the feature list. It is the unit on the invoice. Seats price people. Actions price work. Compute prices time and memory. Each unit creates a different incentive.

Seat pricing: you are buying people, not Agents

Microsoft Agent 365 has a public list price of $15 per user per month. Google Gemini Enterprise starts at $21 per seat per month, with a Standard tier around $30. ChatGPT Business is $20 per user per month on annual billing, or $25 on monthly billing.

Seats are counted by people, not by Agents. Running 100 Agents does not mean buying 100 seats. But the model calls and credits behind each seat are often billed separately.

This is the first hidden layer: a seat is an access right, not a consumption unit.

Seat, action, compute

The same Agent workload looks cheap or expensive depending on the unit you buy.

Action pricing: you are buying work, not answers

Salesforce Agentforce uses Flex Credits: $500 for 100,000 credits. A standard action consumes 20 credits, roughly $0.10. A conversation is priced at $2.

Action pricing turns an "answer" into measurable execution: look up an order, update a record, write a summary. The advantage is alignment with business volume. The disadvantage is that an Agent that tries three times may bill three times.

This is the second hidden layer: trial and error has a price.

Compute pricing: you are buying runtime

AWS Bedrock AgentCore Runtime v2 is priced at $0.1276 per vCPU-hour and $0.0169 per GB-hour. The Gateway is $0.005 per thousand calls.

Compute pricing looks like a cloud server bill: how long it ran, how much memory it held, how many tools it called. It is transparent for engineers, but opaque for business buyers. A business owner cannot easily map a vCPU-hour to a customer outcome.

This is the third hidden layer: engineering efficiency is not the same as business value.

The cheap part is the seat; the expensive part is what is not on the price sheet

Model tokens are often billed separately. Credits, observability, governance, and retries are not on the first line. One request may include planning, three tool calls, two retries, and a human review. The price sheet shows one slice of that work.

This is why two companies can buy the same "Agent OS" and end up with very different unit economics. The difference is not only the vendor. It is the workflow behind the invoice.

The hidden bill

The visible price is a line item. The real cost is a workflow.

The unit determines what you optimize

If you buy seats, you optimize adoption: get more people onto the platform. If you buy actions, you optimize throughput: get more work per credit. If you buy compute, you optimize utilization: keep the runtime busy.

None of these is wrong. But each one can hide the question that matters: who is accountable for the result, who verifies the action, and who pays for the failed attempt?

Three questions before buying

Before comparing prices, compare accountability.

Our approach: account by task, settle by verification

We do not guess cost by seat, and we do not guess value by token. Every task has a start, a verification step, and a failure record. A task that cannot be accounted for cannot be scaled.

Our current stack uses 17 gates, a 50-ticket golden set, and an 82-entry error ledger. That is the source document for the bill: not the model's price list, but the work the system actually performed.

Three actions for your next purchase

  1. Cost one real task. Count the planning step, tool calls, retries, and human review. Do not use a vendor demo as the baseline.
  2. Choose the unit that matches your risk. People are stable, actions fluctuate, compute is an engineering variable.
  3. Add a verification line. If a failed retry is not recorded, the bill will get less transparent as you scale.

Agent OS is not a smarter model. It is a management layer. The seat is the entry ticket; the governance is the operating cost. Buy the seat if you must. Price the governance before you scale.

About the author: Guanlan (观澜) — AI / Agent / digital transformation practitioner. Practical, hands-on writing — follow along and it just works.

Further reading

Top comments (0)