AI pricing models: what to charge for, with real prices from 2026

An AI pricing model is what your customer pays for and how the bill moves as they use more of the product, add more people or get more work done. Seats, tokens, credits, actions, outputs and outcomes are the options, and most AI companies end up combining two of them.

SaaS pricing was built on one assumption: the next user costs you almost nothing. AI broke it. One customer can cost you a hundred times what the next one does, and the more they love the product, the more it costs you to serve them.

A lot of my work at Potio right now is fixing exactly this, usually for SaaS companies that added AI and watched a seat model that worked for years start leaking margin. My position, after a lot of these projects: the model follows from the value metric, most AI companies land on a hybrid, credits are the most common answer and the most often botched, and most of what gets sold as outcome pricing is output pricing.

This guide covers every AI pricing model with real, dated prices from GitHub, Cursor, Salesforce, Intercom and others, how to choose between them, how to design credits that don’t start a revolt, how to protect margin, how to sell to enterprise buyers and what changed in 2026.

What is an AI pricing model?

An AI pricing model has two parts:

• The value metric: the unit you charge for. A seat, a token, a credit, an agent action, a resolved ticket.
• The model: how that unit gets packaged and billed. Flat per unit, a plan with an allowance, prepaid credits, a commitment drawn down over a year.

Why per-seat pricing breaks for AI products

Seat pricing assumes that serving one user costs roughly the same as serving the next. With AI it doesn’t, for two reasons.

Cost now scales with use. Every prompt, agent run and generated image burns inference. A light user and a heavy user on the same $20 seat can be a 100x gap in what they cost you. Unlimited plans make it worse, because the heaviest users are the ones most attracted to “unlimited”. Orb’s Pricing AI Agents report calls this adverse selection: a flat plan fills up with exactly the customers who hurt your margin.

Even OpenAI got caught. A month after launching ChatGPT Pro at $200 a month, Sam Altman posted in January 2025: “Insane thing: We are currently losing money on OpenAI Pro subscriptions!”

Value stops tracking headcount. When the AI does the work, the customer needs fewer people to get the same result. A seat model then charges less exactly as you deliver more.

GitHub Copilot is the public case of both problems. It launched at $10 per user per month, and the Wall Street Journal reported in October 2023 that GitHub was losing over $20 per user per month on average. On 27 April 2026, GitHub announced that from 1 June every chat, agent and code review request would burn AI Credits based on the tokens it used.

Seats still work when value really does sit with a person. At Toggl we priced by seats because every user got their own value from tracking their own time. That logic still holds for plenty of AI products. It just can’t be assumed anymore.

AI pricing models compared

Here are the eight models, roughly in order of where they bill along the customer’s value chain, from inputs (what goes in) to outcomes (what the customer actually gets). Prices are as of October 2026.

Main AI pricing models, with real examples

Per-seat AI pricing

You charge per person with access, often with a fair-use policy in the background.

• Microsoft 365 Copilot: $30 per user a month on an annual commitment for enterprise. Microsoft carries the inference cost inside the seat.
• Salesforce Agentforce add-ons: $125 per user a month for unmetered employee use in Sales, Service and Field Service.
• Vertical AI: legal AI like Harvey sells annual seat contracts to law firms, reportedly at prices far above a typical SaaS seat. The price is anchored to the lawyer’s billable time rather than to inference.

The more vertical your product, the closer to value you can price, because you understand one customer’s economics deeply. A horizontal product has to pick a metric that satisfies everyone, and that metric drifts back toward infrastructure, eventually toward something like electricity. Vertical AI is where seats survive best.

• Fits: value sits with an individual, and usage per person stays within a predictable band.
• Breaks: when a small group of users runs agents all day on the same seat as everyone else.

Subscription tiers with usage limits

A flat monthly plan with an allowance. Go over and you wait, get throttled or upgrade.

• OpenAI: ChatGPT Plus at $20 a month and Pro at $200.
• Anthropic: Claude Pro at $20 and Max at $100 or $200, with rolling usage limits.
• Midjourney: $10 to $120 a month, sold as hours of fast GPU time, with unlimited slower generation on the higher plans.
• Windsurf: in March 2026 it replaced its monthly credit pool with allowances that refresh daily and weekly. Pro is $20, Teams $40 per seat and Max $200.

This is the default for consumer and prosumer AI because people understand a monthly plan. The weakness is the limit itself. Users don’t see it coming and it lands in the middle of the task they care about.

Token-based pricing

You charge per million input and output tokens, usually with discounts for cached prompts and batch jobs. Every foundation model API works this way: OpenAI, Anthropic, Google, and inference hosts like Together AI.

It fits developers who already think in tokens and need to buy raw model capacity. It fails almost everywhere else. A marketing manager can’t tell you whether 400,000 tokens is a lot, and can’t forecast how many she’ll need next month. If your buyer isn’t an engineer, don’t make them price your product in tokens. Translate the work into a unit they recognise.

Credit-based pricing

The customer buys a balance of credits, and each action burns some of them. Credits are usage-based pricing paid up front.

• GitHub Copilot, since June 2026: each plan includes AI Credits equal to its price. Pro at $10 includes $10 of credits, Business at $19 per user includes $19, pooled across the organisation. Code completions stay unlimited. Chat, agent mode and code review burn credits at the token rates of the model you pick.
• Clay: since March 2026 it runs two balances, data credits for marketplace lookups and actions for workflow steps. Launch is $185 a month and Growth $495.
• Salesforce Flex Credits: $500 per 100,000 credits.

There are two kinds. Single-use credits buy one thing each, like an Audible credit buying one audiobook. Multi-use credits are a currency spent across many actions at different rates, which is what most AI products run. More on designing them below.

Per-action and agentic pricing

You charge per task an agent completes. Salesforce prices a standard Agentforce action at 20 Flex Credits, which works out to $0.10. Checking an order status might take two actions.

This fits when actions are discrete and roughly comparable in cost. It breaks when they aren’t. An agent that answers a one-line question and one that refactors a codebase can both count as “an action”, and only one of them costs you real money.

Output-based pricing

You charge per finished piece of work.

• Intercom Fin: from $0.99 per outcome, on top of seats at $29, $85 or $132 a month.
• Zendesk: roughly $1.50 per automated resolution on committed volume and $2.00 pay as you go, according to Intercom’s own pricing guide. Zendesk doesn’t publish a rate card.
• Salesforce Help Agent: $2 per resolution.

Output pricing sits further along the value chain than tokens or actions, which is why it sells. The fight is over what counts as done. Fin, for example, can count a conversation as resolved when the customer leaves without asking for more help, so a customer who gave up in frustration can look exactly like a happy one.

Outcome-based pricing

You charge for a result on the customer’s own books: revenue closed, churn prevented, cost saved. It’s the model every founder asks about on the first call, and the one I’ve almost never seen work. The outcome sits so far down the value chain that you can’t measure it cleanly or prove your product caused it.

Almost everything sold as outcome pricing is output pricing. A resolved ticket is a state change inside the vendor’s software, not cash moving in the buyer’s business. I wrote a full breakdown of why outcome-based pricing is mostly fairy dust and the narrow conditions where it does hold up.

Hybrid pricing: where most AI companies land

A base fee or seats, plus a usage, credit or output layer on top. Intercom is seats plus Fin outcomes. Cursor is a plan that includes a pool of usage, billed at API rates once you go past it. GitHub Copilot is now seats plus credits.

Salesforce is the extreme version. It sells $2 conversations, Flex Credits at $0.10 per action, $125 per-user add-ons and Agentforce 1 Editions from $550 per user a month that bundle 2.5 million Flex Credits a year, all at the same time. That’s one company letting each buyer pick the shape of bill they can live with.

Hybrids win because of a trade-off with three corners, and you only get two:

• Value alignment: the customer pays more as they get more.
• Predictability: finance can budget the line a year out.
• Margin protection: your inference cost is always covered.

The base fee buys predictability. The usage layer buys alignment and protects margin. You can also combine usage with tiers. For a North American automotive software client selling an API on usage, I added a higher tier where the same call returns enhanced data that’s hard to get elsewhere. The price per call is 50% higher, and over 30% of new customers chose it.

How to choose an AI pricing model

A good AI pricing strategy starts from the value metric, not the model. Find the unit that tracks what the customer gets, and it usually tells you which model to use. I’ve written up the full process for finding your value metric, including the filters a good one has to pass.

The rule I use: go as close to the customer’s outcome as you can while the metric is still easy to measure, easy to attribute and something the customer is happy to pay more of. Never trade a strong metric for one that only sounds closer to value.

Then answer four questions.

1. Does your cost per customer vary widely? If your heaviest customers cost ten times your median ones, you need a usage component. A flat price will be wrong for both groups.
2. Does value track a person? If each user gets their own value and usage per person is stable, seats can stay, maybe with a fair-use cap.
3. Can the customer predict and control their usage? If not, a raw meter will scare them. Wrap it in a plan, credits or a commitment.
4. Is your product horizontal or vertical? Horizontal products drift toward infrastructure pricing because one metric has to suit everyone. Vertical products can price much closer to value. One AI infrastructure client of mine did both: a general-purpose model for the wide market, and vertical models (healthcare, for example) that earn more per deal and are harder to undercut.

Then watch how people use the product.

• Bursty usage (a solo user, a prototype, the occasional big job) suits top-ups. You’re selling flexibility.
• Sustained usage (a team using it every day) should push customers up a tier. You’re selling predictability. Selling top-ups to steady users trades good recurring revenue for one-off cash.

How to design AI credits customers trust

Credits are everywhere in AI pricing right now, and they’re behind most of the angry customer threads too.

They sit between a licence, where the vendor gets the cash and the customer carries the risk, and pure usage, where the customer carries little risk and the vendor waits for its money. That gives you a few things other models can’t:

• Customers spend more freely, because the money already feels spent.
• A leftover balance prevents churn. A licence buyer who overpaid downgrades. A credit buyer with credits left keeps using them.
• You can reprice without renegotiating, by changing how many credits an action costs.
• Fast burn is a free upgrade signal. A customer who’s used half their annual credits by month four is telling you something.

That third lever is also how credits go wrong. Three rollouts from the last 18 months show the pattern:

• Cursor, June 2025. On 16 June Cursor moved Pro from 500 requests a month to $20 of usage billed at model API rates. Users couldn’t see how fast they were burning, surprise charges followed, and Cursor refunded unexpected usage between 16 June and 4 July.
• Windsurf, March 2026. It dropped its credit system for daily and weekly allowances, which users found harder to read. It grandfathered existing Pro and Teams prices indefinitely and offered refunds to anyone unhappy after a free extra week.
• GitHub Copilot, June 2026. Promotional allowances for existing Business and Enterprise customers ($30 and $70 a month) cushioned June to August. When they ended, a plan’s credits went back to equalling its price. GitHub also removed the fallback to a cheaper model when credits run out, so work just stops unless an admin allows overage.

None of those were problems with credits as a model. They were problems with visibility and with changing the deal under people. My rules for credit systems:

1. Keep the math simple. $1 = 1 credit, $1 = 100 credits or $10 = 1 credit. Not $1 = 37.
2. Show a live balance and a forecast. “You’ve used 50% in six months, you’re on track” stops most support tickets before they’re written.
3. Give admins spend controls. Caps, alerts and per-team budgets, shipped before the meter goes live.
4. Let credits expire. A 24-month life is a sane default: credits bought in year one roll into year two and expire at its end. Burn the oldest first.
5. Never refund credits as cash. Otherwise you’re running a bank.
6. Don’t silently change what a credit buys. If an action gets more expensive, tell customers before it happens, with the reason.
7. Use a transposed commitment at renewal. What the customer actually spent last period becomes the commitment for the next one. It’s automatic and there’s nothing to argue about.

How to protect margin: price the task, not the token

Falling token prices won’t save your margin. Prices per token have collapsed, with frontier models down more than 80% since March 2023 on BenchLM’s index. But tokens per task have exploded. Stanford’s Digital Economy Lab found agentic coding tasks use roughly 1,000 times the tokens of a simple question. Cost per completed task falls far slower than cost per token, and sometimes rises.

That’s why AI gross margins sit well below SaaS. Bessemer put LLM-first companies at a median of about 65% in 2025 and ICONIQ’s 2026 data came in at a 52% weighted average, against roughly 80% for software overall. I go deeper on this in AI unit economics.

What I see in client data is that a small group does most of the damage. Your total margin gets eaten by 1% to 5% of your customers. So:

• Track gross margin per customer, not just in total. A blended 60% can hide a handful of accounts at minus 200%.
• Deal with the biggest offenders first. Adjust their limits, or change what the expensive actions cost in your credit or usage model. If you run credits, you’ll be tuning what everything costs for as long as the product exists.
• Put margin into sales comp. It’s becoming normal to pay reps partly on deal margin, because it’s easy now to sign a big AI deal that loses money on every unit.

A hypothetical to show the shape of it. You sell a $30 plan to 1,000 customers. The median customer costs you $4 a month in inference. Thirty power users cost $250 each.

• If all 1,000 were median customers, gross profit would be $26,000 on $30,000, about 87%.
• With the thirty power users, it’s $18,620, about 62%.

Thirty customers, 3% of the base, took more than a quarter of your gross profit. An allowance sized to normal use, with overage or a higher tier for the 3%, fixes it without touching the price for anyone else.

Should you charge for failed runs? It depends how close to value you price. If you bill per output, like Intercom’s Fin, charging only for successful work is closer to value and better in the long run. Customers prefer it, and you’ve already priced the failures into the per-unit rate. If you bill for inputs like tokens or compute, the customer chose the work and pays for it whether it succeeds or not. There’s more to a good value metric than this, but it’s a reasonable test.

How to monetize AI features in an existing SaaS product

This is the most common situation I walk into. A SaaS company on seats rebuilds its product around AI, and the seat model stops making sense.

It depends on what your value metric turns out to be. There’s a case for moving fully to usage eventually. But you have something a new AI startup doesn’t: an existing customer base whose behaviour you can see. Model every customer’s bill under the new system before you announce anything. That’s easy to do and almost nobody does it carefully enough.

One client I worked with shows why. They were a North American, sales-led SaaS company on pure seat pricing, moving their product from SaaS to AI, and our target was a full move to credits. The problem was that seats had never been allocated efficiently. Some customers had lots of seats and little usage. Under credits, several of them would have paid less, and on a sales-led base that’s real money walking out the door.

So we ran a transition period with a double ladder. Customers bought credits for usage, paired with seats. I think that’s a fair system for any seat-based SaaS company moving to usage. You protect revenue from accounts that bought for headcount, and you start charging properly for accounts that actually use the AI.

For existing SaaS I go case by case. But more often than not, you’ll spend a while with one foot in SaaS pricing and one in AI pricing.

Pricing AI for high-ticket and enterprise clients

Enterprise buyers want AI and they don’t want a variable bill. Procurement buys predictability, so large contracts wrap usage in commitments: an annual commitment drawn down over the year, lower unit prices for bigger commitments, a true-up when usage runs over, and caps with alerts before a budget is hit. Salesforce’s $125 per-user Agentforce add-ons with unmetered employee use exist for buyers like this. Some buyers need a fixed line on the budget more than they need a fair price per action.

What’s changed in big-ticket AI deals is the size of the first commitment. Customers commit far less up front than they did for SaaS. Low first-year commitments are normal now, and a customer who succeeds can grow that tenfold within the first year. Price for that: make it cheap to start and easy to grow.

The commitment structure I like most for this is a greater-of cumulative commitment. The commitment grows each month, and each month you bill whatever brings total payments to date up to the greater of cumulative commitment or cumulative usage. A quiet month’s slack counts toward the commitment instead of being lost.

Take a $1,000 monthly commitment:

In month 2 the customer went $500 over, but only paid $100 extra because the $400 they didn’t use in month 1 counted toward it. A plain monthly minimum would have billed $1,500. The customer feels treated fairly, and you still collect at least the commitment.

At renewal, use the transposed commitment from the credits section. Last period’s actual spend becomes the next commitment.

AI pricing trends in 2026

The trends in AI model pricing for 2026, each with a date you can check:

• Flat plans moved to metering. Cursor moved to usage-based compute limits in June 2025. Windsurf swapped credits for daily and weekly allowances on 19 March 2026. GitHub Copilot put chat and agents on token-based AI Credits on 1 June 2026.
• Credits became the default meter for AI apps, and the default complaint. The rollouts that went badly all lacked a clear view of the balance.
• Vendors stopped picking one model. Salesforce sells conversations, Flex Credits, per-user add-ons and editions side by side.
• Support AI is settling around $1 to $2 per resolution. Intercom from $0.99, Zendesk around $1.50 to $2.00, Salesforce’s Help Agent at $2.
• Model prices keep falling. Anthropic cut Claude Opus 5.5 from $5 to $4 per million input tokens and $25 to $20 output on 5 October 2026. Agents keep using more tokens per task, so cost per task falls much slower.
• Seats held on in vertical AI and enterprise productivity. Microsoft 365 Copilot is still a $30 seat, and legal AI still sells seats priced against the lawyer’s time.
• Enterprise commitments got smaller up front, with growth expected inside the first year.

Common AI pricing mistakes

• Billing business buyers in tokens. Translate the work into a unit they understand.
• Offering unlimited plans. The heaviest users find them first.
• Repricing credits without telling anyone. It reads as a hidden price increase, because it is one.
• Calling an output an outcome. Buyers notice, and the contract fills up with exclusions.
• Picking a model because it’s trending. Credits spreading through AI changes what customers expect. It doesn’t mean credits fit your product.
• Launching a meter before spend controls. Caps, alerts and a live balance come first.
• Tracking margin only in total. The damage hides in a few accounts.

Choosing your AI pricing model

Pick the value metric first and let it choose the model. For most products that ends in a hybrid: a base that gives the buyer predictability, and a usage or credit layer that covers your inference and grows with the customer.

Of all the AI pricing models, credits are the most flexible and the easiest to get wrong, so give customers a live balance, simple math and controls before you ship. Watch margin customer by customer. For enterprise, wrap usage in a commitment that starts small and can grow.

Frequently asked questions

What is the most common pricing model for AI products?

Hybrid. A platform fee, plan or seat, plus a usage layer, most often credits. Pure pay-as-you-go is mostly limited to model APIs and developer infrastructure, and pure seat pricing is mostly limited to enterprise copilots and vertical AI. GitHub Copilot, Cursor, Intercom and Clay all run some form of base plus usage.

Should an AI product charge per token or per credit?

Per token only if your buyers are developers buying raw model access. Everyone else should get a unit they recognise, and credits are usually the easiest way to do that while keeping your costs covered. Peg credits to simple math, show the balance live and tell customers before you change what a credit buys.

Is outcome-based pricing better for AI agents?

Rarely, because true outcomes, like revenue or retention on the customer’s books, can’t be measured or attributed cleanly. What works is output pricing, such as a resolved ticket or a drafted document, with a platform fee or monthly minimum underneath to absorb failed attempts. Intercom’s Fin at $0.99 per outcome on top of seats is the standard example.

Will falling token prices make AI products cheaper?

Not by much. Token prices for frontier models have fallen more than 80% since 2023, but agents and reasoning models use far more tokens per task, so the cost of completing a task falls much slower. Most AI companies spend the savings on better models and longer agent runs instead of cutting prices.

Should you charge customers for failed AI runs?

If you price on outputs, no. Charging only for successful work sits closer to value, and customers prefer it. Build the cost of failures into the per-output rate. If you price on inputs such as tokens or compute, the customer directs the work and pays for it either way, but give them the visibility and caps to stop a runaway job.