AI Agent Pricing Models Explained: What Vendors Charge

The AI agent pricing models in use today

A SaaS subscription charges the same price whether a customer logs in once a month or a thousand times. AI agent pricing models don't work that way, because one customer's agent can take ten actions and another's ten thousand on the same plan, so vendors have had to rebuild billing around what the agent actually does.

AI agent pricing models are the ways vendors charge for software that takes autonomous action rather than simply granting access. Buyers pay per seat, per token or API call, per completed outcome, or some mix of these, and the choice usually comes down to how easily the vendor can measure and attribute the value a given agent run produced.

Stigg's pricing analysis, last updated 11 June 2026, groups AI pricing into six patterns. Orb's guide to pricing AI agents narrows the enterprise B2B version down to four of them. Lined up together:

Most vendors don't settle on one. Stigg found that companies typically combine two or three of these as they scale, usually pairing a seat or subscription floor with a usage or outcome meter on top.

Examples from real pricing pages

Four vendors show how these models play out with their own published rates.

Salesforce Agentforce runs on Flex Credits, priced at $500 for 100,000 credits. Each agent action consumes 20 credits, working out to $0.10 per action, while a voice action consumes 30 credits, or $0.15. Salesforce also sells a flat $2-per-conversation option, a $5-per-user-per-month Agentforce User License that still draws down Flex Credits, and Agentforce Max Editions starting at $550 per user per month that bundle in 2.75 million Flex Credits a year. Salesforce's own worked examples show a self-service "where is my order" agent handling 20 requests a day costing about $120 a month, and a 100-user case-management deployment handling three cases per user per day costing about $1,800 a month. Unused Flex Credits do not carry over into the next contract term.

Intercom's Fin AI Agent charges $0.99 per outcome on every plan. An outcome counts when the customer confirms the issue is resolved, doesn't ask for further help after Fin responds, or Fin completes a workflow including a handoff to a human. Intercom bills at most once per conversation no matter how many actions Fin takes, and nothing at all if Fin can't answer or the customer asks for a person. Fin Voice costs $1.99 per voice outcome. Isabel Larrow, Head of AI Support at Anthropic, is quoted on Intercom's pricing page saying her team sees around 560,000 resolutions a month through Fin at about a 79% resolution rate.

Sierra, the enterprise agent company founded by Bret Taylor and Clay Bavor, is built around outcome-based pricing. According to IT Convergence's 25 September 2026 analysis, Sierra's customers pay when the agent achieves a specific, valuable outcome rather than for the compute it took to get there.

Anthropic's Claude API is the plainest usage-based example on this list: Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens, Claude Opus 5.5 runs $4 in and $20 out, and Claude Haiku 5.5 runs $0.10 in and $0.50 out for prompts under 100,000 tokens. Stigg's analysis had cited the prior Sonnet 4.6 model at $3 in and $15 out, so the per-token rate for Anthropic's mid-size flagship model dropped between those two releases.

Custom-built agents show the same usage-based logic from the buyer's side. Conti Digital, which builds agents clients own outright, published its own AWS Bedrock metering bill for three production agents handling a retailer's invoicing, social posting, code changes, and record search: $109.71 in their first month of live operation, forecast to rise to $260 the next month as usage grows. The firm separately prices owned builds from $10,000 for an architecture sprint to $25,000 for a production system, against SaaS tools it says run $20 to $500 a month and managed retainers running four to five figures a month, with an owned build and a managed-service equivalent crossing even somewhere between month five and month twenty by its own estimate.

When each model fits, and when it doesn't

The right model depends on whether the vendor can cleanly measure and attribute what the agent did.

Usage-based pricing fits workloads with a clean, countable unit: tokens processed, documents summarized, leads enriched. It fits poorly when a single run can vary wildly in cost, because the buyer bears that variance directly.

Outcome-based pricing fits agents handling customer support, where "resolved" is close to a universal concept that customer support agents like those on AI Agents Listing's customer support category are built around. IT Convergence's analysis names the real friction: defining what counts as a resolution, unpredictable monthly budgets as ticket volume swings, and an incentive for vendors to favor easy cases over hard ones, since cost rises with the work done instead of flattening out the way per-seat software does.

Hybrid and credit-pool pricing give finance teams a predictable floor while still letting usage scale. Salesforce's move from a Conversations-only model to Flex Credits does exactly this: one fungible credit applies across actions, prompts, translations, and voice, instead of a rigid per-conversation unit.

Freemium suits a land-and-expand plan like Salesforce Foundations, which gives teams a $0 way to try the Agentforce Builder and Agent Script tooling before they commit to a metered plan.

What's new

Three dated sources show how fast this is moving. IT Convergence's analysis, published 25 September 2026, traced Sierra and Salesforce Agentforce's shift toward outcome-based and hybrid billing. Stigg's breakdown of six AI pricing models, last updated 11 June 2026, found most vendors now blend two or three patterns rather than running a single one. Conti Digital's post, published 12 August 2026 and updated 16 August 2026, put a real AWS Bedrock metering bill on the table for production agents instead of a rounded estimate, giving a specific number to what "usage-based" costs in practice.

Key takeaways

• AI agent pricing splits into six recognizable patterns: usage-based, outcome-based, hybrid, credit pools, seat-plus-add-on, and freemium, and most vendors blend two or three.
• Salesforce Agentforce's Flex Credits price each agent action at $0.10 ($500 per 100,000 credits), and each Agentforce Voice action at $0.15.
• Intercom's Fin AI Agent charges $0.99 per outcome and bills nothing if Fin can't resolve the issue or the customer asks for a human.
• Anthropic's current Claude API pricing lists Claude Sonnet 5.5 at $2 in and $10 out per million tokens, down from the $3-in/$15-out rate Stigg cited for the prior Sonnet 4.6 model.
• A custom-built agent's actual compute bill can be small: Conti Digital's three production agents cost $109.71 in their first live month, well below the $10,000 to $25,000 it separately charges to build them.

FAQ

What is outcome-based pricing for AI agents?

Outcome-based pricing charges a customer only when the agent achieves a defined result, such as a resolved support ticket, instead of charging for the tokens or API calls it used along the way. Intercom's Fin AI Agent and Sierra both price this way. IT Convergence's analysis notes the tradeoff: defining and auditing a "successful outcome" gets harder as the task gets more complex.

How much does Intercom's Fin AI Agent cost?

Fin AI Agent costs $0.99 per outcome on every Intercom plan, and $1.99 per outcome for Fin Voice. An outcome counts once per conversation, and Intercom charges nothing if Fin fails to resolve the issue or the customer asks for a human.

How much do Salesforce Agentforce Flex Credits cost?

Flex Credits cost $500 per 100,000 credits. A standard Agentforce action consumes 20 credits ($0.10), and an Agentforce Voice action consumes 30 credits ($0.15).

Why don't AI agents use flat per-seat pricing like most SaaS?

Because agent workload doesn't track seat count. One customer's agent might run far more actions than another's on the same plan, so vendors meter tokens, actions, or outcomes directly instead of charging a flat price regardless of use.

How much does it cost to run a custom-built AI agent?

It depends on what "cost" means. Conti Digital's example put the raw model-metering bill for three production agents at $109.71 in their first live month, while the design and build work behind them is priced separately, from a $10,000 architecture sprint to $25,000 for a full production build.

Compare pricing and capabilities across customer support agents, coding agents, and MCP servers on AI Agents Listing.