Agentforce Is $0.10 per Action. Copilot Studio’s Own Example Is 141 Credits.

$0.10 and 141 credits. A standard Agentforce action is 20 Flex Credits. At the $500 per 100,000 credits on the pricing page downloaded November 10, 2026, that is $0.10. A Copilot Studio agent action is 5 credits. At $200 per 25,000 credits, that is $0.04. Microsoft’s own billing example of one follow-up question on a reasoning model is 141 credits on the token meter alone, $1.128 in a pack, and the feature rates sit on top.

The $0.04 line is a real price. It is the price of one named meter. A purchase order that stops there prices a different event from the one the agent will run.

This page uses the USD column only. Salesforce also prints euro, pound, yen, krona, and Australian dollar prices. Every dollar figure below is arithmetic on a multiplier and a credit price those pages print. None of it is an invoice.

A standard action is $0.10. A help resolution is $2.00.

Salesforce’s page defines an action as one function the agent executes: updating a record, summarizing a case, answering a product question, or running a custom prompt or flow. It says each action is metered on its own, and that “Agentforce actions are 20 Flex Credits, while Agentforce Voice actions are 30 Flex Credits.”

The rate card PDF, titled “Working Version Flex Credits Rate Card 11.2026” and marked “Updated as of November 3, 2026,” is the multiplier table. A credit is still $0.005 at the page’s list price. The card says multipliers may be updated.

A help resolution is twenty standard actions. The pricing page also sells customer-facing agents at $2 per conversation. At list price, one conversation and one Help Agent Resolution land on the same dollar. They are different rows. This page did not run a case inside an org to see which row a given job burns.

The same card prices a sandbox help resolution at zero credits and a production one at 400. A test and a live resolution are not the same meter. The card also prices Data 360 and Personalization. Those rows are not in the table.

Max Edition, on the pricing page, includes the Agentforce add-on and 2.75 million Flex Credits per org per year. Divided by 20, that is 137,500 standard actions. Divided by 400, it is 6,875 help resolutions. The page does not print the Max Edition subscription price next to that allowance, so the credits are not a free-standing $13,750 gift. At $0.005 they are worth $13,750 of list-price credit inside whatever the edition costs.

Two other seat prices are on the page. The employee add-on is $125 per user per month and says usage for employees is unmetered. The Agentforce user license is $5 per user per month and says it requires Flex Credits. Foundations is listed at $0 and names the builder, not a production credit pool.

Copilot Studio’s agent action is $0.04. Its reasoning example is 141 credits.

The licensing guide, read November 10, 2026, says all subscription prices are USD estimated retail, per month, billed annually, and subject to change. A capacity pack is $200 per month for 25,000 credits. Unused pack credits do not roll into the next month. Pay-as-you-go is $0.01 per credit, billed in arrears, and the guide says that meter does not start until pack entitlements are exhausted. The marketing page states the same pack price.

Microsoft Learn, last updated October 31, 2026, is the meter table. One interaction can hit several rows. The page’s own example of a tenant-graph reply is 12 credits: 10 for grounding and 2 for the generative answer.

An agent action, in that document, is a step such as a trigger, deep reasoning, or a topic transition on the activity map. Computer-using agents are billed at that same 5-credit rate. They are not included in the Microsoft Copilot user license. The rest of computer use, including what a desktop agent is, is Can an AI agent use your computer?.

Ten thousand agent actions, and nothing else, are $400 in packs (50,000 credits) and $500 pay-as-you-go. The same count of Salesforce standard actions is 200,000 Flex Credits, $1,000. That comparison holds only when each side’s “action” is the event you actually buy.

Four tool calls are the job where $0.04 stays $0.04.

Microsoft prints three daily formulas that do not use a reasoning model. Converting them at the pack price:

The sales formula multiplies the 100 unlicensed users and not the 50 licensed users named in the same paragraph. That matches the no-charge rule for a signed-in Copilot user. Thirty of those support days, if every day matched the average, would be $1,728 in a pack. That is 30 times their daily estimate. It is not a quote.

Four Salesforce standard actions, at $0.10, would be $0.40 against the $0.16 order. For a chain of tool calls and no reasoning model, the Copilot list line is the lower one. The events still are not the same object. Salesforce’s page says one action can be “summarizing a complex case.” Microsoft’s order formula bills four separate action calls and no summary line.

A reasoning model adds 10 credits per 1,000 tokens.

When the agent uses a reasoning model, Learn says the bill is two meters: the feature rate, plus text and generative AI tools (premium) at 10 credits per 1,000 tokens. Cache reads are not billed. Input tokens, output tokens, and tokens written into the prompt cache are billed. Each model call in the turn counts. The charge is not per message.

The worked example is a customer-support agent on Claude Opus 4.7. One follow-up takes two calls. The billed tokens are 4,600 on the plan call and 9,500 on the answer call, 14,100 in total.

14,100 × 10 / 1,000 = 141 credits.

The page says the feature rates for the activities in the turn are added to the 141, and that the token counts are illustrative. A new conversation in the same example, with less already cached, is 461 credits on the premium meter. At the pack price that is $3.688 before the generative-answer line. One Agentforce standard action does not buy that turn. Eleven of them, at $0.10, still come to $1.10, which is short of the $1.128 premium meter.

Agents “powered by the GitHub Copilot harness” are billed on a different page. This note does not reprice that path. A token bill for a model you call yourself, outside either of these credit systems, is the agent cost calculator.

Voice is a minute, and the action meters come inside it.

Copilot Studio’s voice table is per minute, and the classic answer, generative answer, and agent action are marked included in that minute. Classic voice is $0.08 a minute in a pack. GenAI voice is $0.28. Premium GenAI voice is $0.60. Pay-as-you-go is $0.10, $0.35, and $0.75.

Salesforce’s voice row is a voice action, 30 credits, $0.15 at list. The rate card says some customers are billed on Agentforce Voice Minutes instead. This page did not price that second voice meter. Copilot’s $0.08 is a minute with the action meters inside it. Salesforce’s $0.15 is an action. The two rates sit on different units.

Hosted voice products that sell a minute directly, rather than a credit, are a third price list. ElevenLabs’ agent product is the ElevenAgents note. It is not this table.

Prepaid credits expire on different clocks, and 125% stops the agent.

The Salesforce card says Flex Credits must be used before the order end date, and that no rollover is permitted. A year of credits can sit until that date. Microsoft pack credits die at the month’s end even if the contract is annual. Buying a large Copilot pack for a seasonal agent leaves the unused credits on the table every month. Buying a large Flex Credit pool leaves them until the order ends, and then they are gone.

Copilot Studio’s enforcement rule is specific. At 125% of prepaid capacity, for custom agents on the prepaid model, the environment’s agents are blocked. A blocked agent stops responding, including in a conversation already underway. An environment that bills overage to pay-as-you-go is not blocked when the pack runs out. The extra usage goes to the meter. The Salesforce pricing page, in the copy downloaded the same morning, does not print an equivalent cutoff.

Employee-facing Copilot Studio use is no charge when three things are true: the scenario is business-to-employee, the person has a Microsoft Copilot license, and the agent runs as that authenticated user. The guide says fair-use limits apply and Microsoft can change them. Computer-using agents stay on the 5-credit meter anyway.

What to put on the purchase order

Price the row the job will hit, then write that row into the order.

• A record update or a single tool call is the $0.10 Salesforce action or the $0.04 Copilot action. Ten thousand of those, and nothing else, are $1,000 or $400 at list.
• A resolved help case is the $2.00 Salesforce resolution, or the $2.00 conversation SKU. It is not twenty cents.
• A reply grounded in tenant data, with no reasoning model, is Microsoft’s 12-credit example: $0.096 in a pack. That is four tenths of a cent under a Salesforce standard action, and it is a reply, not a write.
• A reasoning-model reply is the premium token meter. Use Microsoft’s 141-credit illustration as a scale, $1.128 before feature rates, and then measure your own instructions and knowledge. Their page says the counts depend on the agent.
• A voice minute on Copilot Studio starts at $0.08 and includes the action meters. A Salesforce voice action is $0.15 and is not a minute.
• If the users already hold Microsoft 365 Copilot and the agent runs as them, the feature table is no charge, computer use excepted. If they do not, the $125 unmetered Salesforce add-on is the other way off the meter, and it is a seat price, not a usage price.

The loop those credits pay for is still a model proposing a step and a harness running a tool. The credit is the invoice for that step. The step’s name on the rate card decides the invoice.