Flex Credits: estimating Agentforce cost per action

Agentforce cost per action is easy to state and easy to underestimate. A standard action costs 20 Flex Credits, or $0.10. A realistic estimate also counts prompts priced by model tier, voice actions at 30 credits, Data 360 processing for manuals and spare parts catalogues, and test runs in the sandbox. Unused credits expire at the order end date. I estimate per conversation type, count every meter separately and buy credits along the go-live plan. In the worked example below, plain text actions make up only 55 percent of the monthly credits.

Flex Credits as the currency for Agentforce

Flex Credits are the currency in which Salesforce bills Agentforce consumption. The Agentforce pricing page1 lists them at $500 or €500 per 100,000 credits. Every billable event draws a fixed number of credits from the Flex Credits rate card2, updated as of 31 August 2026. An action is “a specific function that an AI agent executes on the platform”, for example reading an order from SAP. Whether an agent pays off at all is covered in Agentforce for manufacturers.

Three details matter for planning:

• Credits come from one pool. Actions, prompts, voice and translation draw on the same credits, and the rate card bills Data 360 in Flex Credits too.
• Deterministic steps are billed too. According to the help article on billable usage types3, actions are billed when they run, also when they run deterministically. Moving a warranty rule into Agent Script makes the agent more reliable, not cheaper.
• One pricing model per org. Flex Credits and the $2 per conversation model are not supported in the same org.

There are three buying models, according to Salesforce Ben’s pricing guide4: Pre-Purchase pays the full term upfront, Pre-Commit bills a baseline monthly in arrears with a true-up at term end, PayGo bills what you use. Overage is billed at the contracted rate without a penalty. The Flex Agreement announced in May 20255 also allows converting user licences into Flex Credits and back.

What do prompts cost, and why does the model matter?

Prompts are easy to overlook, because the examples on the pricing page count only actions. According to the help article, prompt usage is counted per call to the LLM gateway, and one prompt covers up to 2,000 tokens of input and output. A call with 6,500 tokens is metered as four prompts. The price per prompt depends on the model tier, and Salesforce’s model support page6 assigns every model to a tier:

Tier	Credits per prompt	Example models
Basic	2	GPT-5 Mini, Gemini 2.5 Flash, Nova Lite
Standard	4	Claude Sonnet 4.6, Claude Sonnet 5, GPT-5, Gemini 2.5 Pro
Advanced	16	Claude Opus 4.8, GPT 5.5, GPT 5.6 Sol
Starter	2	only for bring-your-own models (BYOLLM)

Advanced prompts have been metered since 21 July 2026. The same 6,500-token summary costs 8 credits on a Basic model, 16 on Standard and 64 on Advanced. For a case summary that a service rep reads and corrects, I would start on Standard and move up only if tests show a quality gap.

The help article also names embedded generative features such as Service Replies. I would budget every prompt template and generative feature in the design.

Voice, sandbox and Data 360: where estimates go wrong

Voice

Voice actions cost 30 credits in production and 24 in a sandbox; the voice example on the pricing page counts four per call. The rate card also lists speech to text at 150 credits per hour of audio and text to speech at 6,000 per million characters, metered independently of other usage types. Ask your account executive whether your voice channel triggers them on top of the voice actions. For a machine builder, voice fits a spare parts line that technicians call from the workshop. In a dealer portal, text is cheaper per step.

Sandbox and tests

Actions in sandboxes and scratch orgs cost 16 credits, prompts the full rate. Agent testing runs only in sandboxes7, and Salesforce notes that running tests consumes credits. A test suite that runs after every change is a recurring cost. How many runs make sense is the topic of testing Agentforce agents before go-live.

Data 360

If the agent grounds answers on manuals or spare parts catalogues, Data 360 processes them. The base tier charges 150 credits per MB for unstructured processing and 600 per MB for intelligent processing. Tiers reset monthly per usage type: above 300,000 credits the rate drops in three steps to a fifth, 30 credits per MB from 12.5 million. Queries cost 3 credits per million rows. Sandboxes have no tiers; unstructured processing costs a flat 120 credits per MB there.

Two more points come from Salesforce Ben’s guide to the Data 360 vector database8: hybrid search costs twice as many credits as vector search, and indexes must be rebuilt when new files are added. Plan the monthly update volume, not only the initial load. More in Data 360 costs for AI agents.

Worked example: a spare parts agent for dealers

Assume a machine builder with SAP, Service Cloud, a dealer portal and about 2,000 MB of spare parts catalogues and manuals as PDF files. The agent answers dealers on part numbers, availability and delivery dates, in the portal and by phone. All numbers are assumptions.

Item per month	Assumption	Credits
Text actions	2,000 portal conversations × 3 actions × 20	120,000
Prompts	2,000 summaries of 3,000 tokens = 2 prompts × 4 (Standard)	16,000
Voice actions	400 calls × 4 voice actions × 30	48,000
Speech to text, if charged	400 calls × 4 minutes ≈ 27 hours × 150	4,000
Data 360 updates	200 MB of changed documents × 150	30,000
Total		218,000, about $1,090

On top come set-up and tests. Processing the 2,000 MB once costs 300,000 credits, or $1,500. Ten test runs with 200 conversations each cost 112,000 credits: 9,600 for actions at 16 and 1,600 for prompts per run. A 200 MB sample of the catalogues in the sandbox costs 24,000 credits; the full 2,000 MB there would cost 240,000. Twelve months at full volume plus set-up and tests come to about 3.05 million credits.

• Text actions are only 55 percent of the monthly credits. Multiplying conversations by actions misses almost half.
• The model is a price lever. On an Advanced model the same summaries would cost 64,000 instead of 16,000 credits a month.
• Included pools run out fast. With set-up and tests, this one agent needs more than the 2.75 million credits per org per year the pricing page lists for Max Edition, and six times the 500,000 in Core Edition. See the editions Core, Advanced and Max.

The calculation for an order status agent, compared with conversation pricing, is in what a dealer service agent costs.

How do you estimate Agentforce costs per action?

1. Count actions per conversation type. The pricing page examples give a feel: two actions for an order status check, five for scheduling a field service appointment. Confirm the count with transcripts from sandbox tests.
2. List every LLM call outside the actions, round its tokens up to 2,000-token units and multiply by the model tier.
3. Treat voice separately: 30 credits per voice action, plus speech meters if they apply.
4. Size the documents for Data 360: initial load and monthly changes in MB.
5. Budget the test load: test cases × runs × actions at 16, plus prompts and a document sample.
6. Spread it over the months: pilot, ramp-up, full volume.
7. Keep a reserve for open meters. Headless Platform Interactions, the meter for agentic MCP and API calls, are “TBA” on the rate card; Salesforce Ben reports9 30 days’ notice before metering starts in production orgs. See MCP access as a budget risk.

Four weeks after go-live, compare the estimate with Digital Wallet10. It shows consumption down to agent, action and channel and alerts at thresholds you set. It requires Data 360 to be turned on.

What happens to unused Flex Credits?

They expire. The rate card says Flex Credits must be used before the order end date, with no rollover into the next term.

I would buy the first year along the ramp-up profile: a small pre-purchase for pilot and tests, PayGo or Pre-Commit for the ramp-up, and a larger volume once Digital Wallet shows real numbers. Because overage is billed at the contracted rate without a penalty, a tight first purchase carries little risk. Also note the period of the included edition credits: the Sales Cloud pricing page11 lists the pools per org and per year, while the editions announcement of September 202612 gives no period. Several agents share one pool, so ask how consumption per agent is monitored. These questions are part of my make-or-buy and licence planning.

Sources12

1. 1Agentforce pricing page (opens in a new tab)salesforce.com
2. 2Flex Credits rate card (opens in a new tab)salesforce.com
3. 3help article on billable usage types (opens in a new tab)help.salesforce.com
4. 4Salesforce Ben’s pricing guide (opens in a new tab)salesforceben.com
5. 5Flex Agreement announced in May 2025 (opens in a new tab)salesforce.com
6. 6model support page (opens in a new tab)help.salesforce.com
7. 7runs only in sandboxes (opens in a new tab)developer.salesforce.com
8. 8Salesforce Ben’s guide to the Data 360 vector database (opens in a new tab)salesforceben.com
9. 9Salesforce Ben reports (opens in a new tab)salesforceben.com
10. 10Digital Wallet (opens in a new tab)salesforce.com
11. 11Sales Cloud pricing page (opens in a new tab)salesforce.com
12. 12editions announcement of September 2026 (opens in a new tab)salesforce.com