You’re about to put a price on your AI feature.
Seat, or credits. That’s the whole decision, and it feels binary.
Here’s what the seat-or-credits debate misses: your customers are living the wrong answer right now, out loud, in public.
On June 1, 2026, GitHub flipped every Copilot plan onto a meter.
Thirty days later, the first full bills landed. And the developers who pay for it stopped talking about code.
They started talking about betrayal.
This is the playbook for pricing AI features so you capture the real cost without torching the trust you spent years building. You’ll get what Copilot’s meter shock actually was, why flat-rate pricing is genuinely cracking, the per-outcome pattern that’s winning, and the exact moves to reprice without the cancellations.
What was GitHub’s Copilot meter shock?
On June 1, 2026, GitHub replaced Copilot’s flat monthly requests with token-metered AI Credits, where one credit equals one cent. Plan prices held, but the included usage shrank. Heavy agent sessions now drain a month’s allowance in days, turning a predictable subscription into a running meter nobody had budgeted for.
That’s the whole event, and the timing is why it’s blowing up this week.
The first billing cycle under the new model closed on June 30. So the second wave of posts, the angrier one, is landing right now, written by people looking at a real invoice for the first time.
The numbers are brutal.
Per Digital Applied’s June 2026 cost audit, Copilot Pro includes $10 of credits at $10 a month, and Pro+ includes $39 at $39 a month.
One editor watched 82% of a monthly allowance evaporate on day one. Heavy agent workloads swung 10x to 100x. One developer’s projected month went from $28 to $746.
You can feel the mood in the language. People are calling it an expiring gift card every 30 days. A token-counting stress test.
The headline prices barely moved. What broke was the ability to know what next month costs. A flat fee you can budget became a meter you have to watch, and that swap is what customers are punishing.
GitHub was blunt about why. In its own announcement of the move to usage-based billing, the company said the flat model “is no longer sustainable” because Copilot “now powers far more complex, agentic workflows that consume far more compute.”
Here’s the uncomfortable part for you: they’re not wrong.
Why is flat-rate AI pricing genuinely breaking?
Per-seat pricing assumed a human sat in the seat.
A person can only type so fast, read so much, and send so many prompts in a day. Flat-rate math worked because human appetite was capped. Then agents showed up with no such limit, ran thousands of requests overnight, and the whole model stopped adding up.
This isn’t only a GitHub problem. It’s the entire category repricing at once.
Anthropic split agentic usage out of its subscriptions. OpenAI’s Nick Turley put the death of the old model plainly.
He told investors that “having an unlimited plan is like having an unlimited electricity plan. It just doesn’t make sense,” as documented in The State of AI’s May 2026 report. That same report notes Uber burned through its entire 2026 AI budget by April.
Four months. A whole year’s budget, gone by spring.
So when a founder in the trenches says per-seat pricing is breaking, believe them. The complaint is real, and it’s yours too.
When one customer with an agent does the work of five, a per-seat plan punishes them for being efficient and punishes you for delivering it.
That’s the trap. The marginal cost of one customer can be wildly higher than the next, and a single flat number can’t hold both. The meter exists because the math forced it.
But the math is not what’s making people cancel.
The real enemy isn’t the meter. It’s the surprise.
Read the backlash closely and you’ll notice something. Almost nobody is arguing that AI should be free. They’re arguing about the ambush.
They got hooked on flat-rate, built a year of workflow on it, and then the meter flipped on with no warning they could plan around. That’s not a pricing complaint. It’s a trust complaint, and it’s the one that actually kills accounts.
The data says this is everywhere.
Zylo’s 2026 SaaS Management Index found that 78% of IT leaders hit unexpected charges tied to consumption or AI features in the past year, and 61% of organizations cut projects or initiatives because of unplanned SaaS cost increases. When the surprise bill lands, the response isn’t a support ticket. It’s a budget lockdown.
The teams running Microsoft 365 Copilot said it best when they described their own rollout: the license is fixed, the meter is not, and nobody had budgeted for the meter yet.
A surprise bill isn’t a pricing problem. It’s a broken promise.
And a broken promise is what turns a healthy account into an angry one overnight.
If you’re planning to move your own customers from seats to credits, this is the exact wall you’ll hit. Not the number. The ambush around it.
Your new metered price can be objectively fairer and still lose you the customer. Fairness they can’t predict feels like a trap. Predictability, not price, is what they’re actually buying.
How should you price AI features instead?
Price on an outcome the customer can count before the bill arrives.
That’s the pattern quietly winning across the market, and it has a name: per-resolution, or per-outcome, pricing. You charge when the AI actually completes the job, not for every attempt or every seat. The customer can forecast it because they already know roughly how many jobs they have.
The receipts are strong.
Per SaaStr’s April 2026 analysis, HubSpot moved its Breeze customer agent to $0.50 per resolved conversation on April 14, 2026. HubSpot’s Chief Customer Officer Jon Dick framed it in one line: “You pay when it works, full stop.”
Intercom grew Fin from $1M to over $100M in ARR on $0.99 per resolution. Zendesk charges roughly $1.50 to $2.00 per automated resolution. Sierra built the model from day one and crossed $150M in ARR.
Here’s the emerging shape of the market, as of mid-2026:
| Vendor | AI pricing metric | Rate |
|---|---|---|
| HubSpot (Breeze) | Per resolved conversation | $0.50 |
| Intercom (Fin) | Per resolution | $0.99 |
| Zendesk | Per automated resolution | ~$1.50 to $2.00 |
| Sierra | Per autonomous resolution | Negotiated |
Bottom line: the winners charge for the job done, not the seat filled or the token burned.
Why does this beat raw credits? Because a resolution is a unit your customer already understands. Tokens are the vendor’s cost. A resolved ticket is the customer’s value. Gartner projects AI agents will autonomously resolve 80% of common customer service issues by 2029, so the volume of these units is only going up.
One honest caveat: a variable bill still spooks a CFO, and stacking per-outcome fees on top of seat fees compounds the pain.
The move isn’t “add a meter.” It’s find the one unit of value your customer can predict, and price on that. This is the same value-metric discipline behind every model in our guide to value-based SaaS pricing beyond Goldilocks tiers.
Pick the one thing your customer can tally in their head before the invoice: a resolved ticket, a booked meeting, a shipped feature. If they can’t forecast the unit, they can’t trust the bill.
How do you reprice AI without the backlash?
You’ve picked a fairer model. Now comes the part Copilot fumbled: the rollout. Four moves separate a clean repricing from a trust crisis.
Show the meter before the bill
The deepest wound at Copilot was no receipt. Teams had no per-request breakdown and no clear token count, so the first they saw of the damage was the invoice.
Do the opposite. Put live usage in front of the customer and fire alerts at 50%, 75%, 90%, and 100% of any threshold, the exact discipline Zylo recommends for consumption contracts. Nobody rages at a meter they can watch. They rage at one they find out about.
Grandfather loudly
Protecting your existing customers is table stakes. Telling them, in a headline, is the actual move.
A silent grandfather clause earns you nothing, because the fear spreads faster than the fine print. Say it plainly: your price is locked, here’s for how long, here’s what changes and when. Loud reassurance is cheaper than a win-back campaign after they’ve already panicked.
Cap the downside
Open-ended consumption is where budgets break and churn begins. Cap overage rates, pool commitments across teams so a light month offsets a heavy one, and offer true-down rights, not just true-ups.
A ceiling on the worst case is often what closes the deal, because you’re selling the one thing the meter took away: a known maximum.
If you’re already fighting cancellations, our playbook on reducing SaaS churn covers the save-offer mechanics that pair with this.
Have the willingness-to-pay conversation first
This is the one Copilot skipped, and it’s the one that matters most.
Every pricing disaster starts the same way: a room full of smart people guessing what customers will tolerate, then finding out on the invoice. The fix is to swap the guess for the answer before you flip the meter, not after.
Go and ask your best customers what a fair unit is, what number makes them walk, and where the anger would land. Not a mass email nobody opens. Real conversations, at the depth where people tell you the truth.
That’s the work hollie can do for you. She has real conversations with your customers and brings back what they’ll actually pay and where the trust is thin, ranked.
So you reprice from evidence instead of a hunch. It’s the difference between repricing on data and repricing on hope, and the same instinct behind collecting customer feedback that’s actually usable.
Frequently asked questions
Is per-seat pricing dead for AI products?
Not dead, but breaking where agents replace human work. Per-seat assumes one human of capped output per seat. When an AI agent does the work of five people from one login, seats punish efficient customers and cap your revenue. Expect a hybrid: seats for access, a value metric for the work.
Should I switch my AI feature to credits?
Only if the customer can predict the credit. Raw token credits push your cost accounting onto the buyer, which is what made Copilot’s change feel like a trap. Meter on a unit of value they can forecast, a resolution or a completed job, and show live usage so the bill never surprises them.
What is per-outcome or per-resolution pricing?
You charge only when the AI completes the job, not per attempt or per seat. HubSpot bills $0.50 per resolved conversation, Intercom $0.99, Zendesk roughly $1.50 to $2.00. It ties your revenue to delivered value and gives the customer a unit they can count in advance, so the variable bill feels fair.
How do I raise AI prices without losing customers?
Show usage before the bill, grandfather existing customers out loud, cap overages, and talk to real customers about willingness to pay before you change anything. The backlash Copilot triggered was about surprise, not amount. Remove the surprise and most of the anger goes with it.
Meter honestly. Price the outcome. Talk first.
The Bottom Line
The flat-rate AI era is ending because agents broke the math. But the backlash isn’t about the meter. It’s about the ambush.
Price on an outcome your customer can predict, show usage before the bill, and grandfather loudly. Above all, learn what they’ll actually pay before you flip the switch, not after. Let hollie have those conversations and reprice from evidence, not a guess.
