Usage-based pricing for AI charges customers by what they consume: tokens, credits, units of work such as minutes or pages, or outcomes. It protects margin because heavy users pay their real cost. In 2026 most AI products use credits bundled into subscription tiers, such as Lovable's 100 credits for $25, with published task costs, spending caps and top-ups.
Key takeaways
- AI products moved to usage pricing because cost per user varies hugely. GitHub said a short chat and a multi-hour agent session had cost it the same.
- There are four AI meters: raw tokens, credits, units of work and outcomes. Credits inside subscription tiers are the most common.
- Build your credit rate card from your cost per task type, not a guess. Some tasks make 95% margins, others lose money.
- The four factors from Chapter Four still apply: land and expand, the right usage metric, the prediction problem and sales compensation.
- Prevent bill shock with default caps, pre-task estimates, a 'what things cost' table and a cheaper fallback mode.
In Chapter Four of The Subscription Playbook, I joked that if Netflix used usage-based pricing, its price would change depending on how many shows you watched. "And that's a terrifying thought for me, given how easy it is to get lost in a Netflix binge."
Now imagine the opposite problem.
Imagine you're Netflix, every episode a subscriber watches costs you real money, and some subscribers watch 50 times more than others. Your flat monthly fee would be profitable on the light viewers and a disaster on the bingers.
That's exactly where AI products found themselves.
In April 2026, GitHub explained why it was moving Copilot to usage-based billing. Under its old model, a short chat question and a multi-hour agent session had cost the customer the same. GitHub called that "no longer sustainable". From 1 June 2026 it replaced "premium requests" with dollar-denominated AI Credits, charged at each model's token rate. Plan prices stayed at $10, $39, $19 and $39.
One developer summed up the reaction: "You will get less, but pay the same price."
Both sides of that story matter. GitHub had to fix its margin. Its customers felt short-changed. Usage-based pricing for AI is about doing the first without causing the second. Let me show you how.
Prices checked 27 October 2026. They change often, so see the AI Pricing Index for the latest.
Why Are AI Products Moving to Usage-Based Pricing?
Three reasons.
1. AI has a real cost per use. Traditional software costs almost nothing to serve one more user. AI doesn't. Bessemer found its fastest-growing AI companies averaged about 25% gross margin, "often negative". ICONIQ put the average AI product's gross margin at 45% in 2025, against the 75% to 80% classic SaaS is used to. (The guide to AI gross margins goes deeper.)
2. Usage varies wildly between customers. Agents resend context at every step and reasoning models generate extra "thinking" tokens. Your heaviest users can cost many times your lightest.
3. The market has moved. ICONIQ's 2026 survey found consumption pricing rose from 35% to 42% of AI companies. Among the 44 companies that crossed $100 million ARR in 2026, 18 of the 33 with visible pricing (55%) charge mainly on consumption, as I found in my analysis of 2026's fastest growers. Even the billing companies are following: Stripe completed its acquisition of usage-billing company Metronome on 14 January 2026, with Patrick Collison saying the shift toward usage-based models "will be a defining feature of the next decade".
In the book I called usage-based pricing "the emerging pricing model" and predicted that AI and automation would push more businesses toward it. I didn't expect AI to be the thing being metered. (Life's funny like that.)
The Four Meters for AI Usage Pricing
| Meter | How it works | Real examples (Oct 2026) | Best for |
|---|---|---|---|
| Tokens | Charge per million input and output tokens, often at or near the model vendor's rate | Claude Sonnet 5.5 $2 / $10; v0 publishes its own rates from $0.20 / $1.20 up to $10 / $50; Cursor and Devin bill overage at API list price | Developers who already think in tokens |
| Credits | An abstract currency that each task draws down | Lovable Pro from $25 for 100 credits; Venice and Perplexity price 100 credits at $1; ElevenLabs $99 for 600,000 credits | Apps with many task types |
| Units of work | Per minute, page, request or document | Deepgram from $0.0043 per audio minute; Exa $4 to $12 per 1,000 searches; Firecrawl credits per page | APIs and single-purpose tools |
| Outcomes | Per result achieved | Fin $0.99 per resolution; Zendesk $1.50 to $2.00 | Agents doing countable work |
Most AI products now combine one of these with a subscription: a monthly fee that includes credits, then top-ups. That's a hybrid pricing model, and it's the default for good reason. For a deeper look at the second row, see credit-based pricing. For the first, see token-based pricing. For the last, see outcome-based pricing.
AI Token Cost CalculatorEstimate monthly LLM API spend across models.Open the free tool →The Four Factors, Applied to AI
In Chapter Four I gave four factors for getting usage-based pricing right. Every one of them applies to AI, some more than ever.
Factor #1: Land and expand
Start customers small and let usage grow with them. ElevenLabs gives away 10,000 credits a month free, then steps up from $6 to $990 a month. It reported more than $330 million ARR at the end of 2025. Lovable's cheap entry plans led to its software reaching employees at nearly two-thirds of the Fortune 500, according to the company. (See the ElevenLabs and Lovable case studies.)
Factor #2: Choose the right usage metric
The metric must grow with the customer and communicate value. Tokens grow with your cost. Your customer doesn't care about tokens. They care about messages sent, minutes of audio, pages built, tickets resolved.
So charge in the customer's language and translate it to cost behind the scenes. Lovable does this well: its pricing page shows that "make the button gray" costs 0.50 credits, "add authentication" 1.20 and "build me a landing page" 1.70. Suddenly an abstract credit means something. (The Value Metric Finder helps you score options.)
Factor #3: The prediction conundrum
In the book I wrote about this from the seller's side: usage revenue is harder to forecast. With AI, the bigger problem is often the buyer's side. Customers can't predict their bill.
Replit's effort-based pricing charged on the time and compute each request used. One user reported $1,000 in a week, against $180 to $200 a month before. Replit later added cheaper agent modes, cut cloud prices and launched a free allowance. (The full story is in the Replit case study.)
The fix is in the next section.
Factor #4: Proper sales compensation
If you sell AI to businesses, a $10,000 deal that grows to $100,000 of usage is worth far more than one that stays flat. Pay salespeople on committed usage and on how usage grows over the following year, just like Anna in the book's example, who sold fewer accounts than John but brought in customers that would scale.
Worked Example: Building a Credit Rate Card That Makes Money
Here's where most AI founders go wrong. They set credit prices by feel, then discover some tasks lose money.
Three tasks, one credit price, very different margins
Assumptions (illustrative): you sell 100 credits for $25, so each credit is worth $0.25 of revenue. Token counts below are made-up but realistic; prices are Anthropic's list rates on 27 October 2026.
| Task | Model | Tokens per task | Your cost | Credits charged | Revenue | Margin |
|---|---|---|---|---|---|---|
| Small edit | Haiku 5.5 ($0.10 / $0.50) | 20K in, 1K out | $0.0025 | 0.5 | $0.125 | 98% |
| New feature | Sonnet 5.5 ($2 / $10) | 120K in, 6K out | $0.30 | 1.2 | $0.30 | 0% |
| Full page | Opus 5.5 ($4 / $20) | 200K in, 24K out | $1.28 | 1.7 | $0.425 | −201% |
Fix the full page: route it to Sonnet 5.5 and cache 80% of the input (cache hits cost $0.10 per million on Sonnet). Cost: 160K × $0.10 + 40K × $2 + 24K × $10, all per million = $0.34. At 1.7 credits ($0.425), margin is now about 21%.
Fix the feature: charge 1.5 credits ($0.375) instead of 1.2. Margin 20%.
The lesson isn't the specific numbers. It's that one flat-feeling credit hides a spread of about 300 percentage points in margin across tasks. You have to build the rate card bottom-up, from the cost of each task type, then decide which tasks you'll run cheaply to win customers and which must carry the profit.
And re-check it often. Model prices change. Google's Gemini 4 Argon doubles from $2/$10 to $4/$20 per million tokens on 1 December 2026. If you build on a model with a scheduled rise, price at the post-increase rate.
Model your own tasks with the AI Token Cost Calculator, then check the overall picture in the AI Gross Margin Calculator.
Turn your cost per action into a credit rate card
The free Credit Pack Designer turns your cost per task into credit prices, pack sizes and top-ups with the margin you want.
How Do You Stop Bill Shock?
This is the difference between AI usage pricing that customers love and the kind that ends up on Hacker News.
- Turn spending caps on by default. Supabase ships its Pro plan with the spend cap enabled. Customers can raise it, but they choose to.
- Show an estimate before expensive work. Replit's backlash came from customers seeing the cost only after the agent ran.
- Publish a "what things cost" table, like Lovable's example prompts.
- Use daily or weekly limits, not just monthly. Windsurf (now Devin Desktop) moved from monthly credits to daily and weekly quotas in March 2026, which stops customers running out of a month's allowance in week two.
- Offer a cheaper fallback instead of a wall. Adobe keeps standard generations unlimited after premium credits run out. Midjourney offers an unlimited "Relax" mode on higher tiers.
- Let some unused credits roll over. ElevenLabs allows up to two months, capped at three times the monthly quota.
- Never switch everything off. Netlify's pricing page says that when one project exceeds limits, "all projects on your account will be paused". That's the kind of rule that makes customers shop around.
- Change allowances openly. Most of 2025 and 2026's backlash came from shrinking what a plan included at the same price. If you must, give notice and grandfather existing customers.
When Is Usage-Based Pricing the Wrong Choice for AI?
It's not always the answer.
- Cheap consumer plans. Consumers hate meters. ChatGPT Go is a flat $8 a month and may include ads. Google AI Plus is $4.99. Use flat pricing with fair-use limits at the low end, and meter only the expensive features.
- AI that costs you very little. If AI adds a few cents per user per month, bundle it into the subscription. Metering it adds friction for no benefit. (See adding AI to a SaaS product.)
- Buyers who need a fixed budget. Large enterprises and public-sector buyers often want one annual number. Offer a committed contract or a capped plan.
- When you can't measure it. Chapter Four warned that tracking usage accurately can cost more than it earns for some smaller businesses. If you can't show the customer a clear usage log, don't bill them on usage.
What to Do Next
- List your AI task types and measure tokens per task from real logs.
- Cost each task on your current model and one cheaper model.
- Pick your meter (tokens, credits, units or outcomes) based on who your buyer is.
- Build the rate card bottom-up with a target margin per task, using the Credit Pack Designer.
- Bundle it into tiers with an allowance big enough to reach the first "aha" moment.
- Add guardrails: default caps, estimates, a cost table, a fallback mode and modest rollover.
- Review monthly against model price changes, and before any change, tell customers first.
If you're new to the model, start with the original usage-based pricing guide from Chapter Four, then come back here.
Read the whole book free
The Subscription Playbook as a free ebook and audiobook. Australian business owners can claim a free hardcover, shipped free.
Claim the free hardcover Get the ebookFrequently asked questions
Should I charge customers per token?
Only if your customers are developers who already think in tokens, as with model APIs or tools like v0 that publish per-model token rates. For everyone else, tokens are a cost measure, not a value measure. Convert them into credits or units your customer understands, such as messages, minutes, pages or tasks.
What's the difference between usage-based and consumption-based pricing?
In practice, nothing. Both mean the customer pays for what they use rather than a flat fee. Some people use 'consumption' for infrastructure products, like Snowflake credits or Databricks DBUs, and 'usage' for apps, but the design questions are the same: the unit, the rate, the allowance and the guardrails.
How many credits should my entry plan include?
Enough for a new customer to reach their first real result, the 'aha' moment, without running out. Lovable's guidance is that a small edit costs about half a credit and a whole landing page under two, so 100 credits covers a meaningful project. Test it: if most trial users run out before finishing one job, the allowance is too small.
Can usage-based pricing work for an AI feature inside a cheap consumer app?
It's harder. Consumers dislike meters, which is why ChatGPT Go is a flat $8 a month (and may include ads) and Google AI Plus is $4.99. For low-priced consumer plans, use a flat price with fair-use limits and put metered usage only on the expensive features or higher tiers.
Do credits expire, and should they?
Policies vary. ElevenLabs rolls unused credits over for up to two months, capped at three times the monthly quota. Lovable's monthly plan credits expire two months after issue, while its top-ups last 12 months. Genspark's don't roll over. Some rollover builds goodwill, but unlimited banking creates a liability on your books, so cap it.
Get both books free
The Subscription Playbook (ebook + audiobook) and Feed a Starving Crowd, instantly. In Australia? Claim the free hardcover.
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.
Sources
- GitHub blog, GitHub Copilot is moving to usage-based billing (27 Apr 2026)
- Visual Studio Magazine, Devs sound off on usage-based Copilot pricing (27 Apr 2026)
- SaaStr summary of ICONIQ 2026 State of AI (10 Jul 2026)
- Stripe, Stripe completes Metronome acquisition (14 Jan 2026)
- Lovable pricing (checked 27 Oct 2026)
- Anthropic API pricing (checked 27 Oct 2026)
- Bessemer, State of AI 2025 (13 Aug 2025)
- The Register, Replit Agent 3 pricing (18 Sep 2025)
- Netlify pricing (checked 27 Oct 2026)
Facts and prices checked 27 October 2026. Spotted something out of date? Tell us.