How to price an AI product (2026): API costs, margin maths and the pricing models that work

To price an AI product, work out what one use costs you in model fees, then choose a pricing model that keeps your heaviest user profitable. Unlike most software, every request your customers send costs you money, so a price that works for the average user can lose money on the busiest ones. Know your cost per action, set a target margin, and put a ceiling on every plan: a fair-use cap, credits, usage-based billing, a base fee plus overage, or a price per outcome.

This guide uses API prices checked on 26 October 2026 on each provider's own pricing page, and a worked example you can copy for your own feature. Prices change often, so recheck them before you set yours.

The short answer

Price from your costs up, then sense-check against what customers will pay. The order that works:

1. Measure the tokens one action uses (input and output separately).
2. Multiply by your model's per-million-token prices to get cost per action.
3. Estimate actions per month for light, typical and heavy users.
4. Pick a target gross margin and price so even the heavy user clears it, or caps out before they don't.
5. Choose the pricing model that makes that easy for customers to understand.

Most AI products end up with a hybrid: a monthly fee that includes an allowance, then overage, credits or a hard stop. The rest of this guide shows the numbers behind that choice.

Step 1: work out your cost per action

Your cost per action is the input tokens times the input price, plus the output tokens times the output price. Providers price per million tokens (often written MTok), and output tokens usually cost several times more than input tokens.

Input tokens are everything you send: your system prompt, the user's message, any documents or search results you add, and the conversation so far. Output tokens are what the model writes back. On Google's Gemini models, the pricing page says the output price includes thinking tokens, so a model that reasons before answering bills that reasoning as output.

Count tokens on the exact model you'll ship. Anthropic's pricing page notes that Claude 4.7 and later models use a newer tokenizer that produces approximately 30% more tokens for the same text, so a prompt measured on one model won't cost the same on another. Log real token counts from your first users rather than guessing.

Today's API prices (checked 26 Oct 2026)

These are standard per-million-token prices in US dollars from each provider's own page, for a few current models from each. The last column applies them to the example action used throughout this guide: a support-reply draft with 3,000 input tokens and 400 output tokens.

Sources: Anthropic pricing, OpenAI API pricing (standard tier, short context), Gemini Developer API pricing (paid tier, standard). Google's page was last updated 22 October 2026.

A few things on those pages matter for pricing:

• Scheduled rises. Google lists Gemini 4 Argon at US$4 input and US$20 output from 1 December 2026, and Gemini 3.8 Flash at US$1.50 and US$7.50 from 1 January 2027. The same draft then costs US$0.020 and US$0.0075. Build a buffer into your price for exactly this.
• Batch discounts. Anthropic's Batch API gives a 50% discount on input and output tokens, OpenAI's batch table lists gpt-6.1-sol at US$1 input and US$5 output (half its standard rate), and Google's paid tier lists "Batch API (50% cost reduction)". Batch suits work that can wait, such as overnight summaries, not live replies.
• Caching has a write cost. On Anthropic, a 5-minute cache write costs 1.25x the base input price, and cache reads on Opus 5.5 and Sonnet 5.5 cost 0.05x. Gemini 3.8 Flash adds a storage charge of US$0.50 per million tokens per hour for cached content. Google says Gemini 4 Argon supports implicit caching only.
• Small print changes the maths. Claude Haiku 5.5 charges US$0.50 input and US$2.50 output for prompts over 100,000 tokens, and Anthropic and OpenAI both charge a 10% uplift for regional or US-only processing on newer models.

Worked example: margin maths for a support-reply drafter

Here's the full calculation for one feature, so you can swap in your own numbers. The product drafts replies to customer support tickets for small teams.

Assumptions (yours will differ, so measure them):

• Each draft sends 3,000 input tokens: instructions, an extract from the help centre, and the ticket thread.
• Each draft returns 400 output tokens.
• Model: Claude Sonnet 5.5 at US$2 input and US$10 output per million tokens.
• Light user: 100 drafts a month. Typical: 600. Heavy: 3,000.
• Target: 80% gross margin on AI cost. There's no official benchmark for AI gross margin, so treat 80% as an example and pick your own, remembering hosting, payment fees and support come out of it too.

Cost per draft:

• Input: 3,000 x US$2 / 1,000,000 = US$0.006
• Output: 400 x US$10 / 1,000,000 = US$0.004
• Total: US$0.010 per draft

Price for an 80% margin: price = cost / (1 - 0.80), which is cost x 5.

The flat-plan column is (US$29 - cost) / US$29. The typical user hits the target, but the heavy user costs US$30 and pays US$29, so you lose US$1 on them before hosting, payment fees or support. One heavy user can wipe out the margin from several light ones.

Fix 1: include an allowance, then charge overage. Keep US$29 a month, include 1,000 drafts, and charge US$0.05 per extra draft (five times the US$0.010 cost, so 80% margin on overage).

• User at the cap (1,000 drafts): cost US$10, revenue US$29, margin 65.5%.
• Heavy user (3,000 drafts): revenue US$29 + (2,000 x US$0.05) = US$129, cost US$30, margin US$99 / US$129 = 76.7%.

Fix 2: route easy tickets to a cheaper model. Claude Haiku 5.5 costs 3,000 x US$0.10 / 1,000,000 + 400 x US$0.50 / 1,000,000 = US$0.0003 + US$0.0002 = US$0.0005 per draft. If 70% of tickets are simple enough for Haiku and 30% go to Sonnet, the blended cost is (0.7 x US$0.0005) + (0.3 x US$0.010) = US$0.00335 per draft. The heavy user now costs 3,000 x US$0.00335 = US$10.05 a month.

Fix 3: cache the part of the prompt that repeats. Say 2,000 of the 3,000 input tokens are the same instructions and help-centre extract every time. On a cache hit, that part costs 2,000 x US$0.10 / 1,000,000 = US$0.0002 instead of US$0.004. The draft costs US$0.002 (the other 1,000 input tokens) + US$0.0002 + US$0.004 (output) = US$0.0062. That's only while the cache is warm: the first request pays the write (2,000 x US$2.50 / 1,000,000 = US$0.005 for a 5-minute write), so caching pays off when requests come in close together.

Routing and caching cut cost, but they don't remove the need for a ceiling. They give you room to set a friendlier price.

The five pricing models compared

Each model trades simplicity for customers against protection for you. Stripe's pricing models page lists usage-based pricing as including "fixed fee and overage, pay as you go, and credit burndown", alongside flat rate, per-seat and tiered.

Agents are the hardest case, because one task can take many model calls. If you price agents per task or per outcome, define exactly what counts as billable and cap the tokens an agent can spend on any one task.

Per-seat pricing still works for AI features if every seat has a cap. Without one, a seat is just a flat plan with unlimited exposure.

What real AI products publish (checked 26 Oct 2026)

These are the pricing patterns on each product's own page today, described as published.

• Credits by task size: Lovable. Lovable's pricing page says credits in Default Mode "vary based on task complexity", while Plan Mode costs 1 credit per message. Its examples run from 0.50 credits for making a button grey to 1.70 for a landing page with three generated images. Monthly plan credits expire two months after they're issued, and daily build credits don't roll over.
• Free mode plus dollar credits: Replit. Replit's pricing page lists Core at US$20 a month (US$18 a month billed annually) with "US$20 towards more powerful models" plus up to 30 hours of chat on its Free Mode, and Pro at US$100 a month with US$100 towards its most powerful models.
• Shared usage pool: Cursor. Cursor's pricing page sells paid tiers from US$8 a month with a shared AI usage pool, sized as multiples of the Plus tier (4x on Premium at US$30, 15x on Super at US$100, 50x on Ultra at US$200). Third-party models are billed separately by usage at API list price.
• AI credits with base and flex: GitHub Copilot. GitHub's plans page and billing docs price one AI credit at US$0.01. Copilot Pro (US$10 a month) includes 1,000 base plus 500 flex credits, Pro+ (US$39) includes 3,900 plus 3,100, and Max (US$100) includes 10,000 plus 10,000. GitHub says flex allotments may change, unused credits don't carry over, and code completions don't use credits on paid plans.
• Per outcome: Intercom Fin. Intercom's pricing page lists Fin AI Agent at US$0.99 per outcome. An outcome counts when the customer confirms the issue is resolved, doesn't ask for more help after Fin responds, or Fin completes a workflow, including handoffs. Intercom charges at most once per conversation and refunds the outcome if the conversation is reopened.
• Seats with usage tiers: ChatGPT Business. OpenAI's business pricing page lists a Standard seat at US$20 a month billed annually (US$25 monthly) and a Premium seat at US$100 a month billed annually with "5x more usage than standard". It describes usage as "Unlimited subject to abuse guardrails".

Notice what they have in common: every one defines a unit (a credit, a dollar of usage, an outcome, a seat tier) and puts a limit or a meter on it.

Controls that protect your margin

Pricing sets the ceiling, but these product controls decide how close heavy users get to it.

• Per-user caps. Give every plan a hard number and decide what happens at the limit: block, slow down, switch to a cheaper model, or bill overage. Cap tokens per request too, so one huge document can't cost more than a month's fee.
• Model routing. Send each request to the cheapest model that does the job well, and escalate to a bigger one only when needed. In the example above, routing 70% of drafts to Haiku cut the heavy user's cost from US$30 to US$10.05.
• Caching. Put the parts of your prompt that never change (instructions, reference docs) at the start so they can be cached. Check the write cost and how long the cache lasts on your provider's page.
• Batch for work that can wait. Reports, overnight summaries and bulk imports can go through batch APIs at half price.
• Show usage to customers. A visible meter ("412 of 1,000 drafts used") turns a limit into a feature and makes upgrades feel fair. Lovable, for example, shows the cost of each message in the message history.
• Alerts for you. Stripe's usage alerts can email users who hit a limit, remove access when they exceed a free allowance, or tell your sales team when a self-serve user crosses an upsell threshold.

If you're building the product mostly with AI tools yourself, the same discipline applies to your own tool bills: see AI coding tools for founders. For agents that run parts of your business, see how solo founders use AI agents.

How to bill for usage

You need something that counts usage, rates it, and puts it on an invoice. Stripe covers all three, and a merchant of record can sit in front of it if you sell globally.

Stripe. Stripe's usage-based billing page now leads with Metronome, "a Stripe product", for real-time metering, pricing, billing and reporting. Its basic usage-based billing page says Metronome is "the recommended starting point for all new integrations", including prepaid credits, credit burndown and real-time usage visibility. The older Billing Meters API "remains fully supported for existing integrations".

With Billing Meters, you send meter events and Stripe aggregates them at the end of the billing period. Stripe's recording usage page warns that it processes meter events asynchronously, so upcoming invoices "might not immediately reflect recently received meter events". Don't use Stripe's running total as your in-app usage meter: keep your own count.

Credits in Stripe. Stripe's billing credits can be prepaid or promotional, and they only apply to subscription items with metered prices reported through Meters. Stripe prohibits using them as gift cards or stored value, or for payments to third parties. Credits don't expire unless you set an expiry date.

Merchant of record. If you sell to customers in many countries, a merchant of record handles sales tax and VAT for you, which matters more once invoices vary month to month. Compare the options in merchant of record vs Stripe, and check that whichever you choose supports the usage model you've picked.

Common mistakes

These are the errors that turn a popular AI feature into a loss-maker.

• Pricing from competitors, not costs. Two products at the same price can have very different costs per action. Start with your own maths.
• Unlimited plans with no number. "Unlimited" attracts exactly the users who'll cost the most. Even OpenAI's business plan adds abuse guardrails.
• Using the average user. Margin problems live in the top 5% of users. Model the heavy user explicitly.
• Forgetting output and thinking tokens. Output costs five times input on Claude Sonnet 5.5 and gpt-6.1-sol, and Google counts thinking tokens as output.
• Assuming prices only fall. Google has already scheduled rises for two Gemini models. Leave headroom and recheck provider pages every quarter.
• Counting only AI cost. Hosting, payment fees, refunds and support come out of the same margin.
• Vague outcomes. If you charge per outcome, define it in writing, as Intercom does, so customers don't dispute bills.
• Hiding usage. Customers who can't see their usage get surprised by limits and churn.

Checklist: pricing your AI product

1. Log input and output tokens per action from real usage on the model you'll ship.
2. Calculate cost per action from the provider's current pricing page, and note the date you checked.
3. Estimate monthly actions for light, typical and heavy users (look at the top 5%, not the average).
4. Choose a target gross margin and work out the price per action that hits it.
5. Pick a pricing model: flat with a published cap, credits, usage-based, base plus overage, or per outcome.
6. Set a hard ceiling on every plan, including a per-request token limit.
7. Add model routing and prompt caching, and move non-urgent work to batch.
8. Build an in-app usage meter and alerts at 80% and 100% of the allowance.
9. Choose billing: Stripe (Metronome for new usage integrations) or a merchant of record that supports usage.
10. Review your top 10 users' real cost every month and recheck provider prices every quarter.

Where to go next

If you're still finding your first paying users, read getting your first customers without ads before you spend weeks on pricing pages. If you'd like to swap pricing numbers with other founders building AI products, you can request an invite to Remote Builders: it's free, and a person reads every request.