Classic software pricing assumed the next user costs almost nothing to serve. AI breaks that. Every prompt, agent run or generated asset has a model cost, and your heaviest users can cost more than they pay. The fix is a pricing unit that moves with both value and cost, allowances that protect margin, and price changes customers can see coming.
Why AI pricing is different
Each request consumes model tokens or GPU time, so cost scales with usage rather than with the number of customers. That alone would be manageable if users were alike. They aren’t. A light user might run a handful of requests a week while a power user running agents runs thousands, and under a flat price the light user subsidises the heavy one.
The costs themselves also move. Model prices change, new models cost more or less per task, and agentic features burn far more tokens per job than a simple chat reply.
Bessemer Venture Partners’ State of AI 2025 report shows how this lands in the numbers. Its fastest-scaling AI companies, which it calls “Supernovas”, had “only 25% gross margins” on average, “often trading distribution for profit in the short term”, with margins listed as “~25% (often negative)”. A second group of efficient AI companies, its “Shooting Stars”, averaged around 60%, which the report puts slightly below SaaS peers “due to faster growth and modest model-related costs”. On either figure, the old assumption of near-zero marginal cost is gone.
The margin maths with hypothetical numbers
Take an AI writing assistant sold at $30 per seat per month. Each task (one generation, with retries) costs about $0.04 in model fees, and users fall into three groups.
| User type | Tasks per month | Inference cost | Gross margin on the seat (inference only) |
|---|---|---|---|
| Light | 200 | $8 | 73% |
| Typical | 600 | $24 | 20% |
| Heavy | 2,000 | $80 | −167% |
If a fifth of your seats are heavy users, they can wipe out the margin from everyone else. Worse, margin falls just as the product succeeds, because engaged users are the ones who use it most. That is the main reason so many AI companies have dropped unlimited flat pricing.
Before choosing a model, instrument cost per account. For each customer you want revenue, inference cost and gross margin by month. Without that, every pricing debate is opinion.
The four pricing models
Per seat
A fixed price per user per month, sometimes with an AI add-on per seat. It works when usage per user is fairly even, or when the AI is a modest feature inside a larger product.
It breaks when a few users drive most of the cost. It also breaks when AI agents do the work and fewer human seats are needed: if your product replaces work, seat counts can shrink as it succeeds. Fair-use limits per seat and a higher tier for heavy use keep it alive for longer.
Usage (metered)
You charge per unit consumed, whether tokens, requests, minutes, documents or actions. This suits buyers who are used to metered bills, such as developers and data teams, provided the unit maps to value they recognise.
When the unit is technical and buyers can’t predict the bill, it goes badly. “Tokens” mean little to a marketing manager. Budgets, alerts and spending caps are the protection; the usage-based pricing tactic covers them.
Credits
Customers buy or receive a pool of credits, and different actions use different amounts. Plans usually include a monthly allowance with top-ups. Credits make sense when several features have very different costs and you want one simple unit on the pricing page.
They fail when nobody can tell what a credit buys, or when credit costs change often, at which point credits feel like a hidden price rise. Publish a clear rate per action and show an in-product meter.
Outcome-based
You charge only when the AI achieves a defined result: a resolved ticket, a qualified meeting booked, an invoice processed. The outcome has to be clear, measurable inside your product, and valuable enough to cover the cost of failed attempts. Disputed outcomes, or outcomes that depend on things outside your control, sink it. The defence is a precise, published definition of the outcome.
Choosing between them
| Question | Points towards |
|---|---|
| Does cost per user vary widely across your base (say, more than 5 to 1)? | Usage, credits or outcomes, not pure seats |
| Can buyers name and count the unit without your help? | Usage or outcomes |
| Do you have many features with very different costs? | Credits |
| Is there one clear, attributable result customers pay for? | Outcomes |
| Do buyers need a fixed annual budget? | Seats or platform fee with an included allowance |
Write your value unit down in the customer’s language, then check whether it tracks your cost. The value metric lesson and the value-metric pricing tactic cover how to pick a unit customers associate with value. Most B2B AI products end up with a hybrid: a platform or seat fee that includes an allowance, a published metered rate above it, and clear upgrade prompts at limits.
What companies have changed in public
All prices below are as listed on each company’s site in October 2026. They change often, so check the linked pages before relying on them.
GitHub Copilot: from premium requests to usage-based credits
In April 2026 GitHub announced that from 1 June 2026, premium request units would be replaced by GitHub AI Credits, with usage metered by tokens at each model’s published API rates. The post gives the reason: “a quick chat question and a multi-hour autonomous coding session can cost the user the same amount”, and “the current premium request model is no longer sustainable.” Base plan prices stayed the same, each plan includes credits, and admins gained budget controls. Code completions remain included. GitHub’s plans page now states “1 AI credit = $0.01 USD”.
When per-request pricing hides very large cost differences between requests, GitHub’s answer was to move to a unit that tracks cost while keeping the headline plan price stable and handing buyers controls.
Cursor: a usage change, then an apology
In June 2025 Cursor moved its Pro plan from a limit of 500 requests a month to “$20 of included usage” priced at API rates. In July 2025, in a post titled Clarifying our pricing, the company wrote: “We recognize that we didn’t handle this pricing rollout well, and we’re sorry.” It offered refunds for unexpected charges between 16 June and 4 July 2025, and admitted it had not made clear that “unlimited usage” applied only to its Auto mode. Its current pricing page describes allowances that refresh weekly, top-ups and spending caps.
The economics of the change were defensible. The communication wasn’t. What customers mostly objected to was the surprise.
Intercom Fin: priced per outcome
Intercom’s pricing page lists its Fin AI agent at “$0.99 per Fin outcome”. Its FAQ counts an outcome when the customer confirms the issue is resolved, when they don’t ask for more help after Fin responds, or when Fin completes a workflow, including handoffs. Look at how much work went into defining the outcome. That definition is the product.
Salesforce Agentforce: three models on one page
Salesforce’s Agentforce pricing page offers several units side by side. There’s $2 per conversation. There are Flex Credits at $500 per 100,000 credits, with a standard action using 20 credits, which works out at $0.10 per action. And there are per-user licences and add-ons for employee-facing use. Large vendors often let buyers pick the unit that fits their budget process, at the price of a more complex page.
How to test pricing changes
Pricing tests behave nothing like button-colour tests. Customers talk to each other, and existing customers react differently from new ones. A safer order:
- Model it on real data first. Re-price last quarter’s usage under each candidate model and find out who pays more, who pays less, and what happens to gross margin by segment.
- Test on new customers only. Change the price or packaging for new sign-ups, in one region or for a fixed period, and compare conversion, activation and early usage against the previous cohort. Run the A/B test calculator first to check whether your sign-up volume can detect the difference you care about.
- Watch margin and retention alongside conversion. A cheaper plan that converts better but attracts heavy users can lose money, so track gross margin per cohort for at least two billing cycles.
- Move existing customers last, with notice. Give a clear date, show each customer what their bill would have been under the new model, and offer a cap or a transition period.
- Ship controls with the change: usage meters, alerts at 50%, 80% and 100% of allowance, and hard spending caps. GitHub and Cursor both added or highlighted spend controls alongside their changes.
Common mistakes
- Unlimited plans with no fair-use policy. Someone will find the limit for you.
- Pricing in tokens for non-technical buyers, when documents, conversations, images or actions would mean something to them.
- Changing credit costs silently. If an action now costs more credits, say so, with a date.
- Ignoring model cost changes. Switching to a newer model can raise cost per task sharply overnight, so review cost per action every time you switch or upgrade.
- Bundling AI into every plan without measuring use. Track how much each plan actually consumes before the next renewal cycle.
The first number to pull
Before any of the above, add inference cost per account to your revenue reporting so you can see gross margin by customer and by plan. Then plot usage per seat or per account. If the top 10% use several times the median, flat seat pricing is already at risk, and the hybrid above is where to start drafting.
Questions people ask
Should I charge per seat or per usage for an AI product?
Per-seat pricing is simple to buy and forecast but breaks when usage, and therefore inference cost, varies widely between users. Usage or credit pricing ties revenue to cost but makes bills less predictable for buyers. Many companies combine them: a seat or platform fee that includes an allowance, plus metered usage above it.
What is outcome-based pricing for AI?
Outcome-based pricing charges when the AI achieves a defined result, such as resolving a support conversation, rather than per seat or per token. Intercom, for example, lists its Fin AI agent at $0.99 per outcome on its pricing page at the time of writing. It aligns price with value but needs a definition of the outcome that both sides trust.
How do inference costs affect SaaS gross margins?
Traditional software costs little to serve each extra user. AI features add a variable model cost to every request, so heavy users can cost more than they pay. Bessemer Venture Partners' State of AI 2025 report found its fastest-growing AI companies averaged about 25% gross margins, against about 60% for a group of more efficient AI companies.
How do I change pricing for existing customers without losing them?
Give clear advance notice, explain exactly what changes and why, show customers their own usage so they can predict their bill, and offer spending caps. Cursor's public apology in July 2025, after a poorly communicated pricing change, included refunds for unexpected charges and a promise of clearer notice and usage visibility.
