AI Pricing in 2026: How to Price an AI Product Without Giving Away Your Margin

In June 2025, Cursor had a problem.

Its $20-a-month Pro plan had been the darling of software developers. It included 500 "fast" requests to the best AI models, then unlimited slower ones. The trouble was that a few developers ran long, autonomous coding sessions all day, and each of those sessions cost Cursor real money in model fees.

So on 16 June 2025, Cursor changed Pro to "$20 of frontier model usage per month at API pricing." Users who hadn't set spend limits got surprise bills. On 4 July the company published an apology: "We recognize that we didn't handle this pricing rollout well, and we're sorry." It refunded the unexpected charges.

Here's the twist. Cursor kept growing. It said it passed $1 billion in annualized revenue that November.

So the lesson isn't "don't meter AI." The lesson is that AI changes the maths of a subscription, and you have to design your price around that from day one. In the book I opened the usage-based pricing chapter with Procore's "unlimited" plans and asked: there must be some limits to unlimited, right? With AI, there always are. Every task costs you something.

Let me show you how to price an AI product so your best customers are also your most profitable ones.

Why is AI pricing so different from normal SaaS pricing?

Traditional software has a beautiful property. Once it's built, the next customer costs you almost nothing to serve. That's why classic SaaS runs at 75–80% gross margin and why investors love it.

AI breaks that.

Every time a customer asks your product to write, summarize, search or act, you pay a model vendor for the tokens it reads and writes. A light user might cost you $2 a month. A power user running agents might cost you $200. Same plan, same price.

(Guess which one tells all their friends how much they love your product.)

That's why AI businesses carry much thinner margins:

• Bessemer found its fastest-growing AI companies averaged about 25% gross margin, "often negative," against about 60% for steadier ones.
• ICONIQ's 2026 survey put the average AI-product gross margin at 45% in 2025, projected to reach 53% in 2026.
• Cursor itself was reported at -23% gross margin for the quarter to January 2026, according to Contrary Research.

I see this from both sides. I founded Nagaris, an AI platform for regulated industries. The question is always the same: how do you charge for something whose cost moves every time the customer clicks?

The answer starts with picking the right way to charge.

What are the main AI pricing models?

In 2026, AI companies charge in six main ways. Most of them mix two or three. ICONIQ found the average AI company now blends 1.7 pricing models, with consumption pricing rising from 35% to 42% of companies and outcome pricing from 18% to 23%.

The fast growers confirm the shift. Of the 44 companies that crossed about $100M ARR in 2026 (per ARR Club signals), 18 of the 33 with visible pricing charge mainly on consumption: usage, credits, a usage hybrid or outcomes. Only four rely mainly on seats.

You'll find the full list of 40 real companies, sorted by model, in AI pricing examples. For outcome pricing specifically, read the outcome-based pricing guide.

How do you price an AI product? A seven-step method

This is the method I'd follow if I were launching an AI product tomorrow. It builds on the four factors from Chapter Four of the book (land and expand, the right usage metric, the prediction conundrum and sales compensation), updated for a world where every unit has a cost.

Step 1: Name the job, then pick a value metric

What does your customer actually get? Not "AI." They get a drafted contract, a resolved ticket, a minute of dubbed video, a qualified lead.

That thing is your value metric candidate. In the book I said a good usage metric must be value-based, scalable, flexible, feasible and predictable. For AI, add one more test: does it roughly track your cost? Fin's "resolution" passes. Raw tokens track cost perfectly but fail the value test for most buyers.

If you're unsure, score your options with the value metric guide.

Step 2: Work out your cost per task

You need one number: what does a typical task cost you in model fees?

Cost per task = (input tokens × input price + output tokens × output price) ÷ 1,000,000

A short drafting task that reads 3,000 tokens and writes 500 on Claude Sonnet 5.5 ($2 in, $10 out per million) costs $0.006 + $0.005 = $0.011. An agent that takes 20 steps, rereads a growing context each time and writes its own reasoning can easily cost $1 to $4 per run.

Add your other variable costs too: search APIs, speech-to-text, vector databases and payment fees. The token pricing guide walks through the maths line by line.

Step 3: Map your light, medium and heavy users

Averages lie.

In almost every AI product, a small group of users drives most of the cost. Before you set a price, estimate how many tasks a light, a typical and a heavy user will run each month. If you have beta data, use it. If not, assume the heaviest 5% use 10 times the median (and then check).

Step 4: Choose the structure

Pick the model from the table above that matches your buyer and your cost spread. I'll give you a decision framework further down.

Step 5: Set a margin floor and a value ceiling

Your floor comes from cost. If you want a 70% gross margin on a task that costs $0.011:

Price = Cost ÷ (1 − target gross margin) = $0.011 ÷ 0.30 = $0.037

Your ceiling comes from value. Fin charges $0.99 per resolution because buyers compare it with the cost of a human-handled ticket. Harvey's reported $1,200–$2,000+ per seat per month only makes sense because it's sold against lawyer time, not software budgets.

Price somewhere between the two, as close to the value ceiling as your market allows.

Step 6: Add guardrails

Caps, alerts, slow modes and rollover. These are what stop a heavy user from turning into a loss, and a light user from feeling ripped off. Supabase ships with its spend cap switched on by default. Midjourney keeps an unlimited "Relax" mode so nobody hits a hard wall.

Step 7: Publish the exchange rate

If you sell credits or allowances, show what typical tasks cost. Lovable prints examples on its pricing page: "make the button gray" costs 0.50 credits, "add authentication" 1.20. It's the single cheapest trust-builder in AI pricing.

Worked example: why a $30 flat seat lost money and the hybrid didn't

Let's put real numbers on this. Imagine an AI assistant for small agencies. It does two things: quick drafting tasks and a research agent that runs multi-step jobs.

Notice what happened?

Light users still pay $30. Nobody got a nasty surprise. The people who get the most value pay the most, which is exactly what I argued usage-based pricing should do in the book: "The more you give them, the more you earn."

You can run your own version of this in a couple of minutes.

For the full cost breakdown (hosting, support, payment fees and what "good" looks like), read AI gross margins and inference costs.

How much do AI products cost in 2026?

Here's where real price points sit today. Prices checked 27 October 2026; they change often, so see the AI Pricing Index for the live tracker.

Two things jump out.

First, the price ladder has stretched. Consumer AI now runs from $4.99 to $500 a month. A cheap entry tier widens the funnel while a top tier captures the power users who would otherwise wreck your margin.

Second, usage tiers are now sold as multiples ("5x," "20x") rather than published quantities. That's convenient for the vendor. As you'll see next, it's also where a lot of the anger came from.

Each of these companies has a full pricing teardown, for example Anthropic and OpenAI.

What went wrong? The four AI pricing mistakes of 2025–26

So where did it go wrong?

The companies that lost customer trust in the past 18 months mostly didn't raise their prices. They changed what the price included.

"You will get less, but pay the same price"

That line came from developers reacting to GitHub Copilot's move to dollar-denominated AI credits from 1 June 2026. Cursor's 2025 change followed the same pattern, and so did ChatGPT's Pro 200 plan. According to OpenAI's help center and press coverage, OpenAI cut its allowance for new subscribers on 29 September 2026, without publishing the old or new numbers.

The fix: if you shrink an allowance, publish the old and new numbers, give notice and grandfather existing customers. Windsurf (now Devin Desktop) raised Pro from $15 to $20 in March 2026 and kept existing subscribers at $15.

"I can't see the price until after the work"

Replit's effort-based pricing charged for agent work by time and compute. After its Agent 3 launch, one user reported spending $1,000 in a week against about $200 a month before. The logic matched Replit's costs, but customers couldn't predict the bill. Show an estimate before an agent runs, and never charge for the agent fixing its own mistakes.

"It was free, until it wasn't"

Figma gave every seat AI credits, then enforced the limits from 18 March 2026. Netlify moved new accounts to credits, and its pricing page says that when one project exceeds its limits, every project on the account is paused. Both are defensible on cost. Both caused a backlash. Warn early, show usage before you enforce, and never let a meter take a customer's product offline.

"Up, down, then cut"

Canva raised US Pro from $144 to $180 a year in July 2026, reverted to $144 in September, then cut the free AI allowance from 200 to 20 uses, according to PricingSaaS. Test price changes on new customers first. Reversing a public increase costs credibility.

For the deeper story, read how Cursor broke "unlimited".

Should you charge per seat, per token, per credit or per outcome?

Here's my decision framework. Find the row that best describes your product.

If you're still torn, the Pricing Model Picker asks the questions for you.

How do the book's frameworks apply to AI?

When I wrote The Subscription Playbook, I called usage-based pricing "the natural evolution of the subscription model." AI has made that true faster than I expected. Here's how the book's ideas map across.

Usage-based pricing (Chapter Four). The three benefits still hold: costs scale for customers, retention is higher because customers shrink rather than cancel, and you can serve everyone from a sole trader to an enterprise. So do the downsides: less predictable revenue and more confusion. The book's fix was to bundle usage with a subscription. That is exactly the 2026 hybrid. See usage-based pricing for the full chapter.

Minimum commitment. I suggested turning usage into something closer to a subscription by asking for a minimum commitment. AI vendors do this everywhere. turbopuffer charges a $16 monthly minimum that counts toward usage. Fin sells standalone with a minimum monthly commitment (for example, 50 outcomes). Sierra's contracts reportedly start at about $150,000 a year.

Net dollar retention. Usage pricing is the engine of expansion revenue. When customers succeed, they use more and pay more without a new sale. Track it with the NRR and GRR Calculator.

MOATS. The T in MOATS is Toughness, and one of its elements is a high gross margin. AI products that resell someone else's model start with a thin one. If your product is mostly a prompt on top of an API, read pricing an AI wrapper before you set a price.

A note for Australian businesses

If you're building or reselling AI in Australia, two things matter that the US guides skip.

Your costs are in US dollars, your prices probably aren't. Every model vendor in this guide bills in US$. If you charge in A$, your margin moves with the exchange rate. Build a buffer into your price and review it every quarter.

GST. Under the ATO's rules on imported services and digital products, offshore suppliers charge GST to Australian consumers, while GST-registered businesses generally shouldn't be charged GST on those purchases. When you sell your AI product to Australian customers, your price needs to cover the 10% GST you collect. See GST on subscriptions for the detail. This is general information as at October 2026, so talk to an adviser about your own situation.

Free Subscription Profit Review

Australian business owners can book a free 30-minute review with a BPAG Accountants adviser. An honest conversation about:

• whether a subscription model could work for your business
• how recurring revenue could cover your fixed costs each month
• how it could lift what your business is worth

Book my free review Call 1300 272 422

No cost, no obligation. General information only until we understand your situation.

What to do next

Here's your checklist. Do it in this order.

1. Write down the job your product does and two or three candidate value metrics.
2. Cost one typical task and one heavy task using current token prices, then cost them again at any announced future price.
3. Estimate your light, medium and heavy users and work out what each costs you a month.
4. Pick a structure using the decision table above. For most businesses, that's a hybrid.
5. Set the price between your margin floor and the value ceiling, aiming for at least 60% gross margin after AI costs.
6. Add guardrails: an included allowance, spend caps on by default, usage alerts, and a slow mode or soft limit instead of a hard stop.
7. Publish a "what things cost" table and a clear policy for how and when you'll change allowances.

If you sell credits, the credit-based pricing guide and the Credit Pack Designer will turn your cost per action into a rate card. If you sell agents, read AI agent pricing.

And ask yourself one question before you publish: if your heaviest customer doubled their usage tomorrow, would you celebrate or panic? Price so the answer is "celebrate."

More in this section

Frequently asked questions

Get both books free

The Subscription Playbook (ebook + audiobook) and Feed a Starving Crowd, instantly. In Australia? Claim the free hardcover.

Leave this empty

First name

Email

CountryAustraliaNew ZealandUnited StatesUnited KingdomCanadaOther

What best describes you?Business owner (services)Business owner (retail / e-commerce)SaaS or AI founderHealth or allied health practiceTrades or home servicesAgency or consultantAccountant or adviserEmployee / marketerStudent / other

Yes, I'd like a free 30-minute Subscription Profit Review with a BPAG Accountants adviser (Australian businesses).

Your details go to Subscription Playbook and our sponsor BPAG Accountants Pty Ltd, who may contact you about the book and, if you ticked the box, your review. Unsubscribe any time. Privacy policy.

Send me the free books

Robert Coorey, MBA

Robert is the author of The Subscription Playbook and Feed a Starving Crowd and a Director at BPAG Accountants in Sydney. More about Robert.