Recommended Free Tools
Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
Salesforce AI is not billed through one universal meter. Depending on the feature and contract, Salesforce can charge for agent actions, prompts, conversations, or another business metric; some plans combine per-user licenses with consumption. Agentforce commonly converts actions into Flex Credits, while Einstein generative features count Einstein Requests in token-based increments. Testing, previews, sandboxes, and production use can all be metered, and API traffic has separate request-per-minute, request-per-hour, and timeout limits. The reliable way to forecast cost is to identify the feature’s native meter, apply the current order-form rate card, include non-production activity, and monitor usage in Digital Wallet and the AI Agent Generative AI Usage Data Model.
What Salesforce AI billing models are available?
Salesforce documents three broad commercial models. The model attached to your order form determines what is measured and how an entitlement becomes a charge.
As an Amazon Associate I earn from qualifying purchases.
- Consumption-based pricing: usage is measured in native units such as agent actions, prompts, or conversations and billed as those units are consumed.
- Hybrid licensing: a per-user license supplies access while additional AI consumption is charged against a usage allowance or rate card.
- Business-metrics pricing: the bill is tied to an agreed business measure rather than directly to each technical request.
Agentic usage means an autonomous agent reasons, performs multi-step tasks, and acts independently. Salesforce meters that activity by actions. Embedded, prompt-based capabilities are metered by prompts, so an action and a prompt are not interchangeable units. Your contract and the currently applicable Agentforce or Einstein rate card control the entitlement, price, and any multiplier.
How much do Agentforce actions, voice actions, and conversations cost?
Salesforce’s public 2026 pricing page lists Flex Credits at $500 per 100,000 credits and Conversations at $2 per conversation. The same page lists 20 Flex Credits for a standard Agentforce action and 30 Flex Credits for an Agentforce Voice action. Those public figures are reference prices; procurement should use the rate and entitlements in the applicable order form.
#1 Best Overall
| Metered item | Published amount | What the number means |
|---|---|---|
| Flex Credits | $500 per 100,000 | Public Salesforce 2026 pricing-page rate; equivalent to $0.005 per credit before any contract-specific terms. |
| Standard Agentforce action | 20 Flex Credits | One standard action consumes 20 credits, or $0.10 at the published credit rate. |
| Agentforce Voice action | 30 Flex Credits | One voice action consumes 30 credits, or $0.15 at the published credit rate. |
| Conversation | $2 per conversation | Separate conversation meter listed on the public pricing page; do not assume it is priced like an action. |
Flex Credits are fungible across Actions, Prompts, Translations, and Voice Actions. That shared pool makes the mix of features important: a team that shifts from text actions to voice actions can consume credits at a different rate without changing its total entitlement. Digital Wallet provides visibility into consumption so administrators can compare actual activity with the purchased allowance.
How are Einstein Requests counted?
Einstein Requests pass through the large-language-model gateway and are counted in 2,000-token blocks. Salesforce rounds the combined input and output token total up to the next block, so a request with 6,500 total tokens counts as four prompts, not 3.25 prompts.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsThe October 24, 2025 rate card applies a multiplier to the prompt type. The multiplier is not a claim that every prompt has the same underlying model cost; it is the rate-card factor used for that usage category.
| Prompt category | Multiplier on the effective Oct. 24, 2025 rate card | Qualification |
|---|---|---|
| BYOLLM Starter Prompts | 4 | Bring-your-own-LLM starter category. |
| Salesforce-enabled foundational Basic Prompts | 4 | Basic prompts using Salesforce-enabled foundational models. |
| Standard Prompts | 10 | Standard prompt category. |
| Advanced Prompts | 38 | Advanced prompt category. |
Salesforce says usage types and multipliers can change. For a forecast, preserve the token-volume assumption separately from the multiplier, then recalculate against the rate card that is effective for the contract period rather than carrying the October 2025 factors forward indefinitely.
Which Salesforce AI activity is free, and which activity is metered?
Design work is treated differently from execution. Agentforce Builder design and development features do not consume credits, but running a preview does.
Activities that can be unmetered at the feature level
- Creating and developing an agent in Agentforce Builder, without executing a preview, does not consume credits.
Activities that are metered
- Chat or voice previews launched from Builder are metered.
- Testing through Agentforce Builder, Agentforce Grid, Testing Center, and a sandbox is metered.
- Production interactions are metered under the consumption model attached to the feature.
- Agentforce Grid is metered in every lifecycle phase, not only after deployment.
Therefore, a sandbox is not a billing-free rehearsal environment for AI. Include automated test suites, preview sessions, and Grid runs in the same usage estimate as live traffic unless your contract explicitly supplies a different entitlement.
When can Salesforce AI usage be unmetered?
Unmetered status is conditional, not an inherent property of a feature. Salesforce requires all documented conditions for the particular AI usage type to be met; otherwise the call is metered.
For global unmetering, the documented requirements are the Unmetered User-Based AI permission, execution as the current user, and no automated-process identity. The documented ceiling is 30 LLM calls per minute per user. Salesforce notes that this ceiling is not currently technically enforced, but it remains the stated limit for capacity planning and governance.
Rank #3
- SALESFORCE CERTIFIED ADMINISTRATOR RAPID CERTIFICATION EXAM PREP GUIDE: Quick Prep for Certification Exam Guide for Salesforce Admin Certification with questions and practice tests
- ABIS BOOK
- Independently published
Unmetered refers to the AI call under those conditions, not to every service invoked by the call. Data 360 or another underlying service can still consume its own credits, so an unmetered AI request does not guarantee a zero-cost transaction.
What API limits apply, and what happens when they are exceeded?
Credit availability and API throughput are separate controls. An integration can have unused credits and still receive a limit or timeout error.
| Interface or method | Documented limit | Failure behavior or scope |
|---|---|---|
| Models API, LLM generation | 2,000 requests per minute per org per REST endpoint | Limits can vary by model. An exceeded limit returns HTTP 429 with an error object. |
| Models API, embeddings or feedback | 1,000 requests per minute per org | Separate from generation traffic; an exceeded limit is handled as a rate-limit response. |
| Apex methods in a sandbox | 500 requests per hour | Sandbox-specific hourly allowance. |
| Apex methods in demo or trial orgs | 150 requests per hour | Lower hourly allowance for these org types. |
| Agent API call duration | 120 seconds | A timeout returns HTTP 500, rather than a 429 rate-limit response. |
Salesforce directs customers to their account executive when a Models API limit needs to be increased. Build clients to recognize HTTP 429, apply bounded backoff, and avoid sending an immediate burst that repeats the rejection. Treat HTTP 500 from the Agent API as a timeout path: record the request, decide whether the operation is safe to retry, and prevent duplicate business actions.
Government Cloud customers use api.gov.salesforce.com as the base endpoint. Hard-code the correct host through environment configuration rather than assuming a commercial-cloud endpoint will work for every org.
How should you forecast Salesforce AI spend?
A useful forecast separates the commercial meter from the engineering limits. Work through these steps before setting a budget or volume commitment.
- Classify every feature. Record whether each workflow is an agent action, voice action, prompt, conversation, minutes-based service, or another business metric. Do not convert an action estimate into a prompt estimate without a documented mapping.
- Map the entitlement. Check the order form and the current Agentforce or Einstein rate card for included user licenses, Flex Credit quantity, prompt multipliers, conversation terms, and any geographic or edition qualifications.
- Model token volume. Estimate input plus output tokens per Einstein request, round each request up in 2,000-token blocks, and apply the multiplier for its prompt category. Keep unusually long prompts and generated responses as a separate scenario because rounding makes small requests behave differently from large ones.
- Add non-production traffic. Count Builder previews, Testing Center runs, sandbox tests, and every Grid phase. A forecast that includes production only will understate consumption when a team is iterating rapidly.
- Separate downstream services. Track Data 360 and other underlying-service credits independently, including when an AI call qualifies as unmetered.
- Check throughput. Size worker concurrency and queueing against the per-org, per-endpoint, and org-type request limits. Include a retry policy for 429 responses and a timeout policy for 500 responses.
- Reconcile actuals. Use Digital Wallet and the AI Agent Generative AI Usage Data Model to compare activity by org, feature, user, and lifecycle stage with the assumptions in the forecast.
How can teams troubleshoot an unexpected charge or failed request?
A usage report is higher than production traffic
First separate design activity from executed activity. A Builder configuration may be free while its chat or voice preview is metered, and sandbox, Testing Center, and Grid activity can add consumption before a release reaches production. Use the usage data model to identify which org and lifecycle phase generated the records.
Requests fail even though the credit balance is positive
Inspect the HTTP status and the endpoint. A 429 indicates a throughput limit, not an exhausted Flex Credit pool; compare the request rate with the Models API allowance and ask the account executive about an increase if sustained volume requires it. A 500 from the Agent API points to the 120-second timeout path and should be handled separately from rate limiting.
Prompt costs vary between apparently similar calls
Compare total input and output tokens, then identify the prompt category. Rounding at 2,000-token boundaries and the difference between multipliers of 4, 10, and 38 can change the charge even when the user-visible instruction looks similar.
An unmetered workflow still creates a bill
Verify the executing identity and permission conditions first. If they qualify, inspect Data 360 and other services invoked by the workflow; those services retain their own credit meters.
Frequently Asked Questions
Can a positive Flex Credit balance prevent a 429 response?
No. Flex Credits govern consumption entitlement, while the Models API also enforces request-throughput limits. A client can therefore have credits available and still need backoff, queueing, or an approved limit increase.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Why can a sandbox test affect the same budget as a live agent?
Salesforce meters sandbox testing and several pre-production tools, so those calls draw on the applicable consumption model rather than a permanently free test pool. Budget owners should tag test orgs and review them separately in usage records to see how much of the allowance is non-production.
Does the unmetered permission make an integration unlimited?
It only removes metering when every required condition for that usage type is satisfied. The 30-LLM-calls-per-minute-per-user figure still defines the documented capacity, and downstream services such as Data 360 can continue to consume credits.
Which Salesforce endpoint should Government Cloud integrations call?
Use api.gov.salesforce.com as the base endpoint and keep it in an environment-specific configuration. This prevents a commercial-cloud host from being reused accidentally when deploying the same integration to Government Cloud.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
