Haris OdobasicThere are four main AI pricing models, namely per seat, usage-based (tokens, API calls or actions), credit-based (prepaid units the vendor defines) and outcome-based (per resolved case or completed task). Most vendors now combine two or more into a hybrid, and the model you sign up for decides who carries the cost risk when usage grows.
Seats are predictable but ignore how much AI each person consumes. Usage and credits track consumption but are hard to forecast. Outcomes tie price to value but only work when the result is easy to count. This guide compares the four with published list prices checked on 15 October 2026, shows where credit-based billing hides cost, and ends with a checklist for buyers and a framework for SaaS vendors pricing their own AI.
Why AI pricing is different from SaaS pricing
Classic SaaS costs very little to serve one more user, so a flat price per seat worked. AI is different, because every time a product runs a model, retrieves information or executes a workflow, the vendor pays for compute. Model providers price that compute per million tokens, the units of text a model reads and writes. Anthropic, for example, lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens (Anthropic), and OpenAI publishes its API prices on the same per-million-token basis (OpenAI).
That variable cost changes vendor economics. In the 2026 Growth Unhinged survey, the median target gross margin for AI products was about 50%, and only 12% of companies aimed for SaaS-like margins of 80% or more (Growth Unhinged, May 2026). So vendors pass consumption through to the buyer, in one of four ways.
For buyers the stakes are rising too. In our RevOps in 2026 survey, the median respondent already spends 20% of the go-to-market tech budget on AI, and 49% spend a quarter or more.
The four AI pricing models compared
| Model | What you pay for | Example (published list price, 15 Oct 2026) | Predictability for the buyer | Link to value | Main risk |
|---|---|---|---|---|---|
| Seat | Each named user | Microsoft 365 Copilot at $30 per user per month, paid yearly | High | Weak, because it ignores usage | Idle seats; heavy users squeeze vendor margin |
| Usage | Units consumed, such as tokens, calls or conversations | Salesforce Agentforce Conversations at $2 per conversation | Low | Medium, as it tracks effort rather than results | Bills spike with retries and longer workflows |
| Credit | Prepaid units the vendor converts into actions | Salesforce Flex Credits at $500 per 100,000 credits, with one action using 20 credits ($0.10) | Medium | Depends on what a credit buys | Opaque conversion; unused credits expire |
| Outcome | A completed result | Fin AI Agent at $0.99 per outcome, with a $49 monthly minimum including 50 outcomes | Medium | Strong | Disputes over what counts as an outcome |
| Hybrid | A base fee plus one of the meters above | GitHub Copilot Business at $19 per seat per month including 1,900 AI credits, then $0.01 per credit | Medium | Medium | Two meters to manage |
Seat-based AI pricing
Seat pricing charges per named user per month, with AI bundled into the tier or sold as an add-on. Microsoft 365 Copilot is the best-known add-on, at $30 per user per month paid yearly on top of a qualifying Microsoft 365 licence (Microsoft).
For buyers, seats are easy to budget and approve. The risk is paying for people who rarely use the AI. For vendors, the risk runs the other way. A heavy user can cost far more to serve than a light one at the same price, and an agent that works without a human needs no seat at all. That second point is reshaping SaaS revenue, which we cover in AI vs seat-based pricing.
Usage-based AI pricing
Usage pricing charges for what is consumed, whether that is tokens, API calls, conversations, generated assets or actions. Model providers price this way, and so do many agent products. Salesforce, for example, still lists Agentforce Conversations for customer-facing agents at $2 per conversation (Salesforce).
Usage pricing aligns cost with consumption and lets buyers start small. Its weakness is forecasting, because a workflow that retries, loops or grows longer consumes more without producing more. Our Five Dots RevOps knowledge base notes that unmanaged variance makes customers balk, and the usual remedy is caps or prepaid credits.
Where the hidden cost sits in credit-based AI pricing
You buy an AI tool, pay a subscription fee, and then purchase credits to use its AI features. You can see how quickly the credits disappear, but it is harder to understand what you are paying for. You rarely know which model is doing the work, how much processing a task requires, or whether an unsuccessful attempt costs as much as a useful result. For business leaders trying to manage budgets and measure returns, that leaves an uncomfortable gap between the price of adopting AI and the cost of actually using it.
In the 2026 Growth Unhinged survey, 29% of companies already sell AI credits and another 33% plan to within 6 to 12 months. Three examples show how differently a credit can be defined.
- Salesforce Agentforce Flex Credits cost $500 per 100,000 credits. An Agentforce action uses 20 credits, or 30 for voice, so one action costs $0.10. Unused Flex Credits do not roll over into the next subscription term (Salesforce).
- HubSpot Credits come as a monthly allowance set by your highest subscription tier, for example 3,000 credits on Professional and 5,000 on Enterprise. They pay for usage-based features such as the Customer Agent, Prospecting Agent, Data Agent and Data Studio syncs, expire at the end of each month, and extra credits cost $0.010 each (HubSpot).
- Clay runs two meters. The Launch plan starts at $185 a month, billed monthly, with 2,500 Data Credits and 15,000 Actions a month. Actions reset each billing cycle, and top-ups cost a 30% premium (Clay).
Credit pricing is not inherently unreasonable. Vendors incur costs every time their products run models, retrieve information or execute workflows, and charging for consumption can align price with usage. Our Five Dots knowledge base describes credits as a hybrid in which you pay up front for the right to consume later, which gives both sides some predictability over bursty usage. The problem is that a credit is a unit the vendor defines, often with little connection to an outcome the buyer can evaluate. A team can budget for 10,000 credits without knowing how many completed analyses, resolved cases or usable reports those credits will deliver.
The economics behind the credits
SaaS vendors build products on top of models and translate the underlying token consumption into their own credits. That translation can make purchasing simpler, but it can also hide large differences in what it costs to deliver the same feature.
Take an illustrative example in which a customer pays $100 for credits, and the underlying model usage costs the vendor $30. That leaves $70, or 70% of the credit revenue, before infrastructure, support and other delivery costs. If model selection and workflow improvements cut that usage cost to $10, the remainder rises to $90, or 90%, while the customer's bill stays the same. These are hypothetical figures, not industry benchmarks or product margins, but they show why credit revenue appeals to vendors, and why buyers ask whether efficiency gains ever reach them.
A vendor is entitled to a premium for software that delivers reliable results and saves time. The question is whether the premium reflects differentiated value or a structure that makes comparison difficult. Buyers do not need the vendor's cost base, but they need enough transparency to judge whether spending more produces proportionately better results.
When successful adoption becomes a budget problem
A subscription combined with usage-based pricing can be a sensible model, but it becomes hard to defend when AI is the product's main selling point and meaningful use requires a substantial extra budget. The subscription secures access; the credit bill decides how much value the team can extract.
The exposure is straightforward. If a business budgets $2,000 a month for subscriptions and another $1,000 for credits, doubling consumption takes the monthly bill from $3,000 to $4,000, a 33% increase without adding a single seat. That may be an excellent investment if useful output doubles too. It is harder to justify if the extra consumption comes from retries, longer workflows or experiments that produce little usable work. The metric leadership needs is cost per useful outcome, and credit balances alone do not provide it.
We see the same pattern in client work. Teams worry less about setup fees than about running costs that scale with volume they cannot predict, especially enrichment credits and workflow runs. The teams that keep control meter and optimise rather than abandon the tool. They skip records enriched recently, try cheaper sources first, and monitor consumption before they expand a workflow. Free credits during a trial are useful for experiments, but they prove nothing about unit economics.
Where vendors expose data through APIs or MCP servers, capable buyers can also rebuild parts of a workflow on models directly. Building and maintaining that alternative has its own costs, so it is not automatically cheaper, as our build vs buy guide explains. It does put pressure on vendors to explain what their product adds beyond model access.
Outcome-based AI pricing
Outcome-based pricing charges for a completed result, such as a resolved support conversation, a qualified lead or a finished task. It is the model most aligned with value, and the hardest to get right, because both sides must agree on what counts as a result.
Customer service leads the way because a resolution is countable. Fin's AI agent costs $0.99 per outcome, with a $49 monthly minimum that includes 50 outcomes and no seat costs (Fin). HubSpot moved its Breeze Customer Agent and Prospecting Agent to outcome-based pricing in April 2026, at $0.50 per resolved conversation and $1 per lead recommended for outreach (HubSpot).
It is still a minority model. Only 5% of companies in the 2025 Growth Unhinged survey said outcome-based pricing was their main model, although 25% expected it to be by 2028 (Growth Unhinged, June 2025). Among vendors adding a hybrid AI meter, Bain found only about 10% meter outcomes, against about 35% metering effort and about 55% metering output (Bain, August 2026).
Hybrid AI pricing
A hybrid combines a fixed element, such as seats or a platform fee, with a variable one, such as usage, credits or outcomes. It is now the most common setup, used by 37% of companies in the 2026 Growth Unhinged survey and up from 25% a year earlier.
GitHub Copilot Business shows the pattern with $19 per seat per month, with each seat adding 1,900 AI credits to a shared pool, and usage beyond the pool charged at $0.01 per credit (GitHub). Microsoft sees the same shift across its business applications. On its April 2026 earnings call, Satya Nadella described customers moving "from the traditional seat model to seats plus consumption" and said nearly 60% of its customer service customers already buy usage-based credits (Microsoft).
How SaaS vendors should choose an AI pricing model
If you sell AI, start from the value metric, not from your costs or a competitor's page. Our Five Dots knowledge base uses three tests. The customer understands the unit at a glance, the unit tracks the value they receive, and revenue grows as they get more of that value. A resolved conversation passes all three for a support agent. A token rarely passes the first.
Then cover your variable cost. The same knowledge base notes that AI-native pricing almost always needs a consumption element to cover inference cost, often with credits on top to give customers predictability. If you do not yet know the right metric, a flat early tier that removes seat friction and collects usage data is a sensible start. Learn first and meter later.
Plan the operations before launch. Usage and credit lines must flow through quoting, billing and revenue reporting, which is where many pricing changes stall. Our guide to Salesforce CPQ pricing and the build option covers the quoting choices.
If you are deciding how to price your own AI product, or how far to move away from seats, book a call about our AI pricing work. We model usage, value and willingness to pay, then ship pricing that grows with your customers.
What buyers should ask before they sign
- What exactly is one unit (credit, action, conversation, outcome), and how many units does a typical task use?
- What is included in the subscription, and do unused units roll over or expire?
- What is the overage rate, and can we set hard spending caps?
- Do failed attempts, retries and escalations consume units?
- Can we see usage by user, workflow and outcome, so we can calculate cost per useful outcome?
- How much notice do we get before prices or credit conversion rates change?
- If we build part of this ourselves later, can we export our data?
Inside your own team, meter AI like any other cost. Our Five Dots knowledge base recommends one key and one budget per use case for every model your automations call, so one runaway workflow cannot spend the month's allowance, and comparing each workflow's run cost with the manual cost of the same work. Our AI playbook for sales and RevOps shows how to tie that cost to revenue metrics.
Why predictable AI pricing is a competitive advantage
I do not expect usage pricing to disappear, and unlimited AI at a fixed price will not suit every workload. But vendors can offer more certainty through fixed fees with defined allowances, transparent overages, enforceable spending caps, or pricing tied to completed tasks. Predictable pricing also gives vendors a reason to manage delivery costs, since the right model, cached work and fewer retries protect margin. When every inefficiency simply consumes more customer credits, that incentive is weaker.
The AI pricing model that wins will make adoption easier to approve and expansion easier to justify. Business leaders can accept a healthy vendor margin; what they struggle to accept is an unpredictable bill with an unclear relationship to results.
Whether you are pricing AI or buying it, we can help you get to cost per useful outcome. Book a call.
Frequently asked questions about AI pricing models
What are the main AI pricing models?
The main AI pricing models are per seat, usage-based, credit-based and outcome-based. Most vendors now combine a fixed element, such as seats or a platform fee, with one of the variable meters, which is called hybrid pricing.
What is credit-based pricing?
Credit-based pricing means you prepay for vendor-defined credits and spend them as you use AI features, so the cost of a task depends on the vendor's conversion rate.
What is outcome-based pricing?
Outcome-based pricing charges only when the AI delivers a defined result, such as a resolved support conversation. Fin charges $0.99 per outcome and HubSpot charges $0.50 per resolved conversation for its Customer Agent.
Which AI pricing model is best for buyers?
The best model is the one that lets you forecast cost per useful outcome. Outcome-based pricing aligns best with value; usage and credits work if you get caps, allowances and clear unit definitions.
How do you forecast AI costs under usage or credit pricing?
Measure how many units a typical task uses during a pilot, multiply by expected volume, and add a margin for retries and growth. Then set spending caps and review cost per useful outcome monthly.
