How to price an AI app: credits, tokens, seats or a flat fee?

AI app pricing from unit economics: cost per request, margin targets, heavy users, then credits vs tokens vs seats vs flat fees, with a worked margin table.

Short answer

Price an AI app from the cost of a finished job and from your heaviest user. Most apps fit a subscription with a credit allowance: the base fee covers typical use, and a cap with paid top-ups covers heavy users. Set the cap so a user who hits it still leaves margin, then test prices on profit per visitor.

  • Model choice is the first pricing decision: as of November 2026, the same 5,000-in, 1,000-out request costs $0.001 on Claude Haiku 5.5 and $0.04 on Claude Opus 5.5.
  • Bessemer's 2025 State of AI put fast-growing AI Supernovas at about 25% gross margin and Shooting Stars at about 60%.
  • Double the model cost and the example $29 plan's blended margin drops from 78.5% to 60.9%, while the user at the 600-job cap goes from $9.86 of profit to an $8.14 loss.
  • When every sale carries model cost, the price with the best revenue per visitor can lose on profit per visitor, so test AI plans on a profit goal.

Price an AI app around its heaviest user. At Anthropic's November 2026 rates, one request to Claude Sonnet 5.5 with 5,000 input tokens and 1,000 output tokens costs 2 cents, and a finished job that takes 1.5 requests costs 3. A customer who runs 3,000 jobs a month on a flat $29 plan costs you $90 in model calls and puts you $62 in the red. The fix is a pricing model that caps what one customer can cost you and charges for usage past the cap.

What does one AI request cost you?

Input tokens times the input price, plus output tokens times the output price, plus whatever you pay for cached tokens, tool calls and retries. Model providers publish the per-token rates; your logs supply the token counts.

Prices from Anthropic's pricing page as of November 2026, applied to one request with 5,000 input and 1,000 output tokens:

ModelInput, per 1M tokensOutput, per 1M tokensOne request1,000 requests
Claude Haiku 5.5$0.10$0.50$0.001$1
Claude Sonnet 5.5$2$10$0.02$20
Claude Opus 5.5$4$20$0.04$40

The Haiku 5.5 prices apply to prompts up to 100,000 tokens; longer prompts cost $0.50 and $2.50 per million. The same request spans a 40x range depending on the model, so the model you pick is your first pricing decision, made before any price tag.

Two levers lower the bill without changing the model. Prompt caching bills a repeated prompt prefix at a fraction of the input price: on Sonnet 5.5 cache hits cost $0.10 per million tokens, so if 3,000 of the 5,000 input tokens are a cached prefix, the request drops from $0.020 to $0.0143. The Batch API takes 50% off input and output for work that can wait.

Then convert requests into jobs. A finished job (an email drafted, a document summarized, an image the user keeps) often takes several requests: retries, regenerations, agent steps. In the worked example below, a job averages 1.5 Sonnet 5.5 requests, so it costs $0.03.

What gross margin should an AI app target?

Aim for a blended gross margin of 60% or better at the plan level, and know that the market sits lower. Bessemer's State of AI 2025 found that the fastest-growing AI companies, which it calls Supernovas, averaged about 25% gross margin, while its Shooting Stars averaged about 60%, slightly below SaaS peers. ICONIQ's 2026 State of AI report, based on surveys of over 300 executives at software companies building AI products, puts AI product gross margins at 45% in 2025, with 53% projected for 2026 and 59% for 2027.

Two rules turn that into pricing:

  1. The blended margin across all users meets your target.
  2. A user who hits the plan's cap still leaves a positive margin. If the user at the cap loses money, your most engaged customers are your worst accounts.

Then stress the model cost. Prices usually fall with new models, but they can also rise on a date you don't control: Google lists Gemini 3.8 Flash at $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026, and $1.50 and $7.50 from January 1, 2027 (Gemini API pricing). Rerun your margin table at 2x model cost before you publish a price.

Why do heavy users break flat AI pricing?

Because cost follows usage, and usage is lopsided: a small share of users runs the product all day. Two public cases show what happens next.

In July 2025, Anthropic announced weekly rate limits for Claude Pro and Max subscribers, effective August 28, 2025. It estimated they would apply to less than 5% of subscribers and aimed them at people running Claude Code continuously in the background, around the clock; Max subscribers could buy extra usage at standard API rates (TechCrunch).

In June 2025, Cursor moved its Pro plan from 500 requests a month to $20 of frontier model usage at API pricing. In a July 4 post it explained that the hardest requests cost an order of magnitude more than simple ones, apologized for communicating the change badly, and offered refunds for unexpected usage charges between June 16 and July 4 (Cursor).

The lesson for a small team: decide the limit and the overage path before launch and state them on the pricing page in units your users understand. Then you never have to take something away from a customer who is already paying for it.

Which pricing model fits an AI app?

Usually a hybrid. ICONIQ found that companies blend 1.7 pricing models on average, and that the share using consumption-based pricing rose from 35% to 42% in six months (ICONIQ). These are the five building blocks, with examples as of November 2026:

ModelHow it worksFits whenBreaks whenExample
Flat fee with fair-use limitsOne price, usage limited per time windowMost users stay well under the limitLimits are vague or power users run it nonstopClaude: Pro $20 a month ($17 a month billed yearly), Max $100 for 5x Pro usage, limits per 5 hours and per week
CreditsPrepaid units, in a monthly allowance or packsEach action has a natural unit with a stable costThe unit is abstract and buyers can't predict spendElevenLabs: Starter $6 a month for 30,000 credits, 1 credit per character of text to speech
Usage-based meteringBilled per unit after the factBuyers budget per unit and can map it to valueBuyers fear a surprise billIntercom Fin: $0.99 per outcome, charged once per conversation
SeatsPrice per user, AI includedUsage per person has a natural ceilingOne seat runs automated workloadsNotion: Business plan, priced per member, includes Notion Agent and AI Meeting Notes
Base fee plus creditsA subscription includes an allowance; usage past it costs extraYou need predictable revenue and a ceiling on costThe allowance is too small to feel generousGitHub Copilot: Pro $10 a month includes $15 of monthly AI credits at $0.01 per credit

Flat fee with fair-use limits

The simplest plan to sell and the hardest to keep profitable. It works when the limit is generous for nearly everyone and specific for the few who hit it. Claude's paid plans show the shape: limits apply per 5-hour window and per week, Max at $100 offers 5x Pro usage, and Ultra at $200 drops the 5-hour limit and offers 10x Pro usage per week. Paid plans can also buy usage credits.

If you sell flat, write the limit in a unit the user recognizes (messages, minutes, documents) and decide what happens at the limit: wait for the reset, upgrade, or buy more. Flat plans suit consumer and prosumer assistants, where buyers won't think in units.

Credits

Credits turn a variable cost into a prepaid one. Define one credit as something the user sees: a character of speech, an image, a minute of audio, a processed document. ElevenLabs sells monthly allowances (Creator is $22 a month for 121,000 credits) and lets unused credits roll over for up to two months, capped at twice the monthly quota, while the subscription stays active.

Credits suit generation products where each output has a natural size. The risk is abstraction: if users can't predict how many credits a task takes, they hoard credits or leave.

Usage-based metering

Bill per unit after use: per token, per API call, per resolved conversation. Intercom charges $0.99 per outcome for its Fin agent and counts at most one per conversation. Metering ties revenue to cost exactly, and suits developer platforms and B2B automation where a buyer can map the unit to value.

Consumers dislike open-ended bills, so pair metering with budgets and spend alerts. The trade-offs are in our entry on usage-based pricing.

Seats

Price per user and include the AI. Notion folds Notion Agent and AI Meeting Notes into its per-member Business plan, and sells Custom Agents separately at $10 per 1,000 monthly Notion credits. Seats work when AI is one feature among many and usage per person has a natural ceiling.

Seats break when one seat runs automated jobs around the clock; a usage limit or a credit add-on, like Notion's Custom Agents credits, covers that case. More in seat-based pricing.

Base fee plus credits

A subscription with an included allowance and paid usage past it. GitHub Copilot Pro costs $10 a month, includes $15 of monthly AI credits at $0.01 each, and keeps code completions unlimited on every paid plan: cheap actions stay unlimited, expensive ones draw down credits.

Cursor's plans add a detail worth copying. The included allowance refreshes weekly, so one heavy day can't burn a whole month, and users who run out can buy top-offs (Cursor pricing). For AI apps sold to individuals and small teams, this is the model to start from.

How do you set credit prices and allowances?

Start from cost per credit, set the allowance from real usage, and price top-ups higher than in-plan credits:

  1. Define the credit and keep a cost table per action. In the example, one job is one credit and costs $0.03.
  2. Set the allowance from the usage of active users, generous enough that typical users never think about it.
  3. Check the margin at the cap. A $29 plan capped at 600 jobs leaves $9.86 (34%) on a user who uses all of it. To keep at least 50% at the cap, the cap would be 445 jobs.
  4. Price top-ups for full use. Bought credits get used, so price them at your target margin: at 75%, a job that costs $0.03 sells for $0.12, and a 250-job top-up for $30.
  5. Decide rollover and expiry, and say so on the pricing page.

In-plan credits can be cheaper per unit than top-ups because many subscribers never use their whole allowance. The plan's price absorbs that unused share; top-ups don't have one.

What do the margins look like in a worked example?

Example: a $29-a-month plan with 600 jobs included, $0.03 of model cost per job (Claude Sonnet 5.5, 1.5 requests per job), and Stripe's US card fee of 2.9% + 30¢, which is $1.14 on $29 (Stripe pricing, as of November 2026).

UserJobs a monthModel costPayment feeGross profitMargin
Light60$1.80$1.14$26.0689.9%
Typical200$6.00$1.14$21.8675.4%
At the cap600$18.00$1.14$9.8634.0%
Power user on an unlimited plan3,000$90.00$1.14−$62.14−214%

Margin = (price − payment fee − model cost) ÷ price, before hosting and support. With half the users light, 40% typical and 10% at the cap, average model cost is $5.10 and the blended margin is 78.5%.

Now double the model cost, as Google's scheduled change would for a team on Gemini 3.8 Flash. The blended margin falls to 60.9%, the typical user to 54.7%, and the user at the cap loses $8.14 a month. The cap that looked safe at today's prices needs to come down, or the price needs to go up. The payment fee calculator swaps in your own provider's fees.

How do you test the price of an AI plan or credit pack?

Test the price on new visitors and judge it by profit per visitor. Every AI sale carries model cost, so a cheap price that wins more sales also pays for more usage, and revenue per visitor alone overrates it.

Example, with $5.10 of model cost per sale and the same Stripe fees:

PriceConversionRevenue per visitorProfit per saleProfit per visitor
$194.2%$0.80$13.05$0.55
$293.0%$0.87$22.76$0.68
$392.2%$0.86$32.47$0.71

Revenue per visitor picks $29. Profit per visitor picks $39, because each extra buyer at $29 brings $5.10 of model cost along.

Leverprice supports this directly with a profit goal (Leverprice for AI apps): you set your provider's fee and a unit cost on the product through the API or an AI agent over MCP, and it ranks prices by profit per visitor after both. For the unit cost, enter the expected model cost of what one sale buys: the first month's usage for a monthly plan, or the expected usage of the whole pack for a credit pack. Include today's price as one of the test prices so it serves as the control. The engine moves traffic toward the most profitable price and calls a winner once one price is at least 95% likely to be best (how the engine works).

Illustration: revenue per visitor peaks at a middle price. A fixed cost per sale moves the profit peak to a higher price.

Allowance size is a separate experiment, because it changes what the buyer gets rather than what they pay. Leverprice tests prices; an allowance test needs your own flag. Hold the price, give a random half of new signups the larger allowance with a feature flag (the PostHog price-testing guide shows the flag setup), and compare conversion, the share of users who hit the cap, top-up revenue and model cost per user over at least one billing cycle. Then price-test the allowance that wins. The sample size calculator tells you how many visitors each test needs.

A pricing checklist for an AI app

  1. Log tokens per action and compute the cost of a finished job on your current model.
  2. Pick a unit your users understand and build a credit table from it.
  3. Set price and allowance so the blended margin meets your target and the user at the cap still clears a positive margin.
  4. Rerun the table at 2x model cost.
  5. Define what happens at the cap: wait for the reset, upgrade, or buy top-ups priced for full use.
  6. Publish the limits on the pricing page in units, before anyone pays.
  7. Test the plan or pack price on new visitors with a profit goal and today's price as the control.

Frequently asked questions

Should I show token counts to customers?

Only to developers who already think in tokens. Everyone else needs a unit tied to output, such as images, minutes, documents or credits, with a table that says what each action costs.

What should happen to unused credits?

Pick a rule and publish it. ElevenLabs rolls unused credits over for up to two months, capped at twice the monthly quota, while the subscription stays active, and forfeits them on downgrade or cancellation. Expiry protects your margin; rollover takes away the reason to burn credits before the month ends.

What unit cost should I enter for a profit goal?

The expected model cost of what one sale buys, at realistic usage. In the example that is $5.10, the blended first-month cost, rather than $18 for a user at the cap or $1.80 for a light user.

Can I offer unlimited AI usage?

For cheap actions, yes: GitHub keeps code completions unlimited on paid Copilot plans while other AI usage draws down credits. For expensive actions, "unlimited" needs a written fair-use limit; in the example, one power user on an unlimited plan erases the profit of almost three typical users.

How often should I revisit AI pricing?

Every time you change models or your provider changes its rates. Rerun the margin table each time, because a cap that was safe at last quarter's token prices can lose money at next quarter's.