In February 2014, about two days before Slack started charging money, Stewart Butterfield sat his team down to check the billing system one last time. They used their biggest customer as the test case: WalmartLabs.
According to Business Insider, WalmartLabs had 800 people signed up. Half of them, 400, were inactive. People from other parts of Walmart had joined because their email address matched. Under a normal per-seat model, WalmartLabs would pay for all 800, or someone would have to clean up the list by hand.
So Slack changed the rule before launch. You pay only for people who actually use it. The early version counted a member as active if they had logged in and sent a message in the last 14 days, according to the Building Slack history site. Today Slack’s help page says a member becomes inactive after 28 days without using it, and Slack adds a prorated credit for the unused time.
Butterfield later said the policy earned goodwill and made customers more likely to renew. The interesting part is not the kindness. It’s that Slack changed what it charged per. Not “seats you bought,” but “people who used it.” That one choice is the most important pricing decision a software founder makes, and most of us make it by copying someone else.
Below: what to charge per, which pricing shape fits, how to pick the number, tiers, trials, annual plans, price raises, and testing with early customers. A fictional API runs through the middle.
The value metric: what you charge per
A value metric is the unit your price is attached to. Per user, per message, per document, per gigabyte, per project. When a customer gets more value, the number on this unit should go up, and so should the bill.
A good value metric passes four tests:
- It grows with the customer’s value. A company that sends ten times more messages through you gets roughly ten times more out of you.
- The customer understands it and can predict it. They can guess next month’s bill before it arrives.
- You can measure it cleanly. No arguments about what counted.
- Your costs move with it, or at least don’t run ahead of it. If each unit costs you money to serve, charging per unit protects your margin.
Slack’s original problem was test 1. Seats you bought didn’t track value. Active people did.
Think of value metrics as a ladder, from loosely tied to value to tightly tied:
| Rung | Example | Tied to value? |
|---|---|---|
| Flat fee | One price, unlimited use | Loosely. Heavy and light users pay the same |
| Per seat bought | Every account on the list | Somewhat. More people usually means more value |
| Per active user | Slack’s fair billing | Closer. Only people who use it |
| Per unit of work | Per message, per document, per token | Tightly. Every unit is something the customer asked for |
| Per outcome | Per successful match, per resolved ticket | Tightest. But hard to measure and easy to argue about |
Higher up the ladder is not automatically better. Tighter metrics are fairer, but harder to predict. A finance manager may prefer a flat $200 a month over “probably $150 to $260” even if the second is cheaper on average. Your job is to find the rung that your buyer trusts.
Five pricing shapes, with real examples
Once you have a metric, you wrap it in a shape. Here are the five you’ll see most. All prices are from each company’s official pricing page at the time of writing (November 2026), in US dollars.
1. Per seat
You pay a fixed price for each user, every month. Slack Pro is $8.75 per user per month billed monthly, or $7.25 billed yearly. Its fair billing policy makes the seat “active user” rather than “invited user.”
2. Usage (pay as you go)
You pay for exactly what you use. Most APIs work this way.
- Anthropic charges Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens. A token is a small chunk of text, roughly part of a word. It halves those prices for work sent through its Batch API, which runs requests asynchronously instead of right away.
- OpenAI lists gpt-6.2-sol at $2 per million input tokens and $10 per million output tokens on its standard tier, with a separate, roughly half-price batch tier.
- Twilio charges $0.0083 per text message segment sent or received in the US, plus carrier fees that depend on the recipient’s phone network.
3. Credits
The customer buys a bundle of credits, and different actions use different amounts. It’s usage pricing with a wrapper. ElevenLabs, the voice generation company, sells plans like Creator at $22 a month with 121,000 credits, where one character of text-to-speech generally costs one credit, depending on the model.
4. Tiers (packages)
You sell a few fixed bundles, each with limits. Basecamp is a clean example. Its plans run from Free up to Pro Unlimited at $199 a month billed yearly, or $329 billed monthly, with “unlimited users, no per-user fees.” Smaller plans cap projects, storage and users.
5. Hybrid
A base fee that includes some usage, plus a price for each unit above it. This is the most common shape for small API businesses, because it gives the customer a predictable floor and gives you protection when someone uses ten times their plan. The worked example below uses it.
| Shape | Bill easy to predict? | Protects you from heavy users? | Watch out for |
|---|---|---|---|
| Per seat | Yes | Only if usage tracks people | Shared logins; price stops growing when customers automate |
| Usage | No | Yes | Buyers who hate surprise bills |
| Credits | Partly | Yes | Customers who can’t work out what a credit buys |
| Tiers | Yes | Up to the limits | The limits are your real value metric, so choose them carefully |
| Hybrid | Mostly | Yes | More moving parts to explain |
Price from value, but cost sets the floor
There are three numbers that bound every price.
The floor is your cost to serve one unit. Below it, every sale loses money. For an API that calls an AI model, this includes the model tokens, hosting, storage, payment fees, retries, and a fair share of support. If you haven’t worked this out, start with Unit Economics in Plain English.
The ceiling is the value to the customer. That’s what the work would cost them another way: a person’s time, an existing tool, or a mistake they avoid. Nobody pays more than the value for long.
The competitor band is what similar products charge. Customers will compare. You don’t have to sit inside the band, but if you’re outside it, you need a clear reason.
Technical founders often start from the floor: costs to four decimal places, plus a margin. But for most software, the floor and the ceiling are very far apart. Cost-plus pricing puts you right next to the floor and hands almost all the value to the customer. In my opinion, it’s the most common pricing error among developers.
The rule I’d use: price from the value side, check the competitor band, and use cost only to make sure you’re comfortably above the floor.
A worked example: Ledgerly, an invoice-reading API
Ledgerly is a fictional company. Its API takes a PDF invoice and returns the supplier, date, line items and totals as structured data. Customers are small accounting firms and bookkeeping tools.
Step 1: the value metric
Nobody logs in to an API, so seats make no sense. Per invoice processed passes all four tests: it grows with value, a firm knows roughly how many invoices it handles, it’s easy to count, and each invoice costs Ledgerly money.
Step 2: the floor
Ledgerly sends each invoice to an AI model priced at $2 per million input tokens and $10 per million output tokens (the standard rate on the two real price lists above). An average invoice uses about 3,000 input tokens and 500 output tokens.
| Cost per invoice | Calculation | Amount |
|---|---|---|
| Model input | 3,000 x $2 / 1,000,000 | $0.006 |
| Model output | 500 x $10 / 1,000,000 | $0.005 |
| Text extraction, hosting, storage | Estimate | $0.004 |
| Subtotal | $0.015 | |
| Plus 10% for retries and failed files | $0.015 x 1.10 | $0.0165 |
The floor is about 1.7 cents per invoice. Ledgerly also budgets $3 per customer per month for support, and card fees of 2.9% plus $0.30 per payment.
Step 3: the ceiling
Without Ledgerly, a bookkeeper types each invoice in by hand: say 4 minutes, at $20 an hour.
Value per invoice = 4 / 60 x $20 = $1.33
The ceiling is $1.33 per invoice, about 80 times the floor. A firm handling 400 invoices a month spends about $533 of staff time on it.
Step 4: the band and the plans
Ledgerly’s fictional competitors charge between $0.08 and $0.25 per invoice. Ledgerly goes for a hybrid shape with three tiers:
| Plan | Price per month | Invoices included | Effective price per invoice | Each extra invoice |
|---|---|---|---|---|
| Starter | $29 | 300 | $0.097 | $0.12 |
| Growth | $99 | 1,500 | $0.066 | $0.09 |
| Scale | $299 | 6,000 | $0.050 | $0.07 |
Step 5: check the margins
If a customer uses its whole allowance, here’s what’s left per month after model and hosting costs, card fees and support:
| Plan | Price | Usage cost | Card fee | Support | Left over | Margin |
|---|---|---|---|---|---|---|
| Starter | $29 | $4.95 | $1.14 | $3.00 | $19.91 | 68.7% |
| Growth | $99 | $24.75 | $3.17 | $3.00 | $68.08 | 68.8% |
| Scale | $299 | $99.00 | $8.97 | $3.00 | $188.03 | 62.9% |
These are the worst cases, because most customers don’t use their full allowance. Overage invoices keep 76% to 86% of their price after usage costs.
And the customer side: the firm with 400 invoices a month would fit Growth at $99, against $533 of staff time. It gets about $5.39 of value for every dollar. That’s an easy yes, and Ledgerly still has room to raise prices later.
One more check: a Starter customer who processes 883 invoices pays $29 plus 583 x $0.12, about $99. Above that, Growth is cheaper, so heavy Starter users have a natural reason to upgrade.
Good, better, best and anchoring
Ledgerly’s three plans follow a pattern called good-better-best: a small plan, a middle plan most people should buy, and a big plan.
Why three? First, people with different budgets get a fit. One price either loses the small buyer or undercharges the big one.
Second, the big plan changes how the middle one looks. This is anchoring: the first or largest number someone sees becomes their reference point. Next to $299, $99 feels moderate. Without the $299 option, $99 is the most expensive thing on the page.
A few rules that help:
- Design the middle plan first. It should fit your most common customer. The other two exist around it.
- Make each step up feel worth it. Ledgerly gives five times the invoices for about 3.4 times the price on the step from Starter to Growth.
- Keep the top plan real. If nobody can ever use it, buyers will notice the trick. Some customers should actually buy it.
Free plan, free trial or reverse trial?
You’ll need some way for people to try before they pay. There are three common ones.
| Option | How it works | Good when | Watch out for |
|---|---|---|---|
| Free plan | A limited version is free forever | The product spreads through use, and a free user costs you almost nothing | Free users who never upgrade, and a free plan so good nobody needs to |
| Free trial | Full product, free for a set time, then pay or lose it | The value shows up within days | Trials that end before setup is finished |
| Reverse trial | Full product for a set time, then drop to a free plan instead of losing access | You want people to feel the paid features but stay around | It’s more work to build both versions well |
Slack and ElevenLabs have free plans. Basecamp pairs a small free plan with a 30-day trial of its paid plans.
For an API, there’s a simple version of this: give a starter amount of free usage. My own CV scoring API at cv.phamphutai.com works this way. Its value metric is pay per CV scored, and the first key comes with 100 free CVs, which is enough for a developer to wire it up and judge the results. Pricing beyond that isn’t decided yet, so it’s here only to show the shape.
Whatever you pick, the free part has a cost, so count it. Free users still use your servers and your model tokens. Put a limit on free usage that you can afford if ten times as many people sign up as you expect.
Annual discounts
An annual plan means the customer pays for 12 months upfront, usually at a discount. You get cash now and a customer who is less likely to leave this year. They get a lower price.
How big are the discounts in real life?
- Slack Pro: $7.25 a month billed yearly against $8.75 billed monthly, about 17% off.
- Basecamp Pro Unlimited: $199 a month billed yearly against $329 billed monthly, about 40% off.
A common, simple choice is “two months free”: charge for 10 months, give 12. That’s a 16.7% discount. For Ledgerly, Growth on a yearly plan would be $990 a year, the same as $82.50 a month.
Three cautions:
- Offer it once people are getting value. A yearly plan someone regrets becomes a refund fight.
- Annual cash is not revenue yet. It’s money you owe in service. Don’t spend it as if it’s profit, especially if you’re watching runway, as in Default Alive or Default Dead.
Raising prices and grandfathering
Your first price will be wrong. Usually it will be too low. So plan to raise it.
Grandfathering means keeping existing customers on their old price when new customers pay more. You can do it forever, for a fixed time, or not at all.
Slack gives a real example. In July 2022 it announced its first price increase since launching in 2014: Pro went from $8 to $8.75 per user per month billed monthly, and from $6.67 to $7.25 billed yearly, starting September 1, 2022. According to TechCrunch, existing customers could keep the old price for another year by renewing before the change. That’s a mild form of grandfathering: notice, a clear date and a way to lock in.
Why is raising prices less risky than it feels? Go back to Ledgerly. Say it has 200 Starter customers at $29, each leaving $19.91 a month. If Ledgerly moves Starter to $35, each customer leaves $25.73.
200 x $19.91 = $3,982 a month before
Ledgerly could lose 45 of its 200 Starter customers, about 23%, before monthly profit falls. If it loses 20 customers, 10%, it ends up at 180 x $25.73 = $4,632 a month, more than before.
Not every raise works, but the fear is usually bigger than the risk. More on that in Why Small Service Businesses Charge Too Little.
A simple way to do it: try the new price on new customers first, give existing customers at least 30 days’ notice, say honestly why, and offer a way to keep the old price for a while.
How to test prices with early customers
With 10 or 20 customers, you can’t run a clean pricing experiment. You can still learn a lot by talking.
Talk about their current cost
Before you say any number, ask how they solve the problem today: how long it takes, who does it, what tools they pay for. That’s your ceiling. It’s the same habit as in Your First 10 Customers: learn their world before you pitch yours.
Say a real price, then be quiet
Then say a price, and watch. If they agree instantly, it’s probably too low. If they wince but keep talking about how to get budget, you’re close. If they change the subject, it’s too high or the value isn’t clear yet.
The strongest signal is money. A paid pilot, a deposit or a signed order means far more than “yes, I’d pay that.” See Validate an Idea in One Week for ways to ask for commitment early.
The Van Westendorp questions
A Dutch economist, Peter van Westendorp, published a method in 1976 called the Price Sensitivity Meter. It asks four open questions:
- At what price would this be so cheap you’d doubt its quality?
- At what price would it be a bargain, a great buy for the money?
- At what price would it start to feel expensive, though you’d still consider it?
- At what price would it be too expensive to consider?
With hundreds of answers, researchers plot four curves and read an acceptable price range from where they cross.
Be honest about what this can and can’t do. With a dozen early customers, you won’t get reliable curves. People also answer differently when it’s a survey than when it’s their own card. And the method doesn’t show competitors’ prices, so answers float free of the real market. Pricing researchers also point out that the crossing points have little economic theory behind them.
So use the questions as a conversation guide, not a calculator. If most people say “too expensive” starts around $100 a month and “a bargain” sits near $30, you have a sensible range. Then test real prices inside it, with real payments.
Afterwards, split your retention chart by plan, as in Cohort Retention: The Chart That Tells the Truth. If cheap plans churn fast, a low price may be attracting the wrong customers.
Key ideas
- The value metric is the biggest pricing decision. Slack changed from charging for seats to charging for active users, and customers trusted it more for it.
- Pick a shape your buyer can predict: per seat, usage, credits, tiers, or a hybrid of a base fee plus usage, which suits most small APIs.
- Price from value, check competitors, and use cost as the floor. Cost-plus pricing for software usually leaves most of the value on the table.
- Good-better-best works because of fit and anchoring. Design the middle plan first, and make the top one real.
- Free plans, trials and reverse trials all cost money. Cap free usage at a level you can afford.
- Price raises are usually less risky than they feel. Give notice, start with new customers, and offer a way to lock in.
- Test prices by talking and by asking for money. Van Westendorp’s questions are a useful guide, not a precise answer.
Sources: Eugene Kim, A small change Slack made 48 hours before launch shows why users love its $2.8 billion app (Business Insider, July 31, 2015); Johnny Rodgers and Ali Rayl, Preparing for Launch (Building Slack, April 2024); Slack, Fair Billing Policy and Pricing (November 2026); Anthropic, Pricing (November 2026); OpenAI, API pricing (November 2026); Twilio, SMS pricing in the United States (November 2026); ElevenLabs, Pricing (November 2026); Basecamp, Pricing (November 2026); Slack is increasing prices and changing the way its free plan works (TechCrunch, July 18, 2022); Van Westendorp’s Price Sensitivity Meter (Wikipedia, citing the 1976 ESOMAR Congress paper).
Related: Unit Economics in Plain English, Positioning: Who It’s For and Why It’s Better, and Distribution Beats Product.



