In June 2025 Replit replaced its flat $0.25 per checkpoint with effort-based pricing, charging each AI Agent request on the time and compute it used. Simple edits got cheaper, big tasks cost more, and customers couldn't see the price until afterwards. Some bills jumped. Replit added cheaper modes, cut prices and kept growing. Lesson: estimate first, cap by default, never bill for failures.
Key takeaways
- Effort-based pricing matches price to cost, but only feels fair if customers can predict it.
- Charging after the work is done, including for the agent's own mistakes, breaks trust fast.
- Replit's fixes: cheaper agent modes, model routing, big cloud price cuts and a Free Mode allowance.
- Growth continued: $400 million raised at a $9 billion valuation in March 2026.
- For any variable job, quote a range first and get approval above it.
Imagine calling a plumber about a dripping tap.
He says, "I charge for effort." He doesn't say how much. He gets to work. Two hours later he hands you a bill. Part of it covers the half hour he spent fixing a pipe he'd accidentally loosened himself.
Would you call him again?
That, in a nutshell, is what some Replit customers felt in 2025.
What Changed When Replit Moved to Effort-Based Pricing?
Replit's AI Agent builds and deploys apps from a prompt. Until mid-2025 it charged a flat $0.25 per "checkpoint," a unit of completed agent work.
That was simple, but not fair. A tiny edit cost $0.25. A big task could trigger lots of checkpoints.
On 18 June 2025 Replit introduced "Effort-Based Pricing." Its reasoning made sense:
"The effort involved in addressing user requests can vary significantly from request to request."
So simple requests would usually cost less than $0.25. Complex tasks would be bundled into a single checkpoint that "may cost more than $0.25 but reflects the total effort involved." Users could also pick a "high power model" or "extended thinking" for extra cost.
On paper, it's exactly what I'd recommend: price tracks the work.
In practice, customers couldn't see the price until after the work was done.
What Went Wrong?
Here's the timeline, using reported figures.
| Date | What happened |
|---|---|
| 18 Jun 2025 | Effort-based pricing replaces $0.25 checkpoints |
| 11 Jul 2025 | A billing error overcharges about 6% of users; Replit refunds them plus $10 credits |
| 10 Sep 2025 | Agent 3 launches; The Register reports one user spending $1,000 in a week against $180 to $200 a month before |
| 24 Feb 2026 | New $100 Pro plan; Core cut from $25 to $20 |
| 11 Mar 2026 | Raises $400 million at a $9 billion valuation |
| 1 Aug 2026 | Autoscale compute prices cut 81%, requests 67%, database storage 77% |
| 18 Aug 2026 | "Free Mode" allowance added inside paid subscriptions |
The complaints fell into three buckets:
- No price before the work. You only knew what a request cost after it ran.
- Paying for failures. Users said they were charged when the agent looped or got things wrong.
- Surprises from autonomy. Replit admitted in a recap post that the rollout surprised people, for example when its action-oriented agent changed code when users only meant to chat.
Replit also conceded, according to one pricing analyst, that effort pricing "can end up being more expensive over the lifetime of a project."
Why Did Replit Need to Change Its Pricing at All?
Because it was losing money on heavy users.
Sacra, a research firm, estimated Replit's gross margin fell from 36% to -14% in early 2025, before the repricing, with a 2025 blended figure of about 23%. Treat those as estimates. But they explain the urgency. A flat $0.25 per checkpoint couldn't cover a long, expensive agent run.
Feast or famine, from the other side
Years ago I was the government national sales manager for a telecommunications company with usage-based pricing. One month, a government customer made a crazy number of conference calls during an election. We celebrated. The next month their spending dropped off a cliff and we panicked. Usage pricing makes the seller's revenue lumpy. Replit's story shows it can make the customer's bill lumpy too, and customers like that even less.
Why Effort Pricing Felt Unfair (Even Though It Was Logical)
In the book, I warned that usage-based pricing "is more complicated to market." You can tell customers you'll charge them a tiny amount per use, but they still can't work out what they'll spend.
That's what I call the prediction conundrum. With effort pricing, it lands on the customer.
Compare the three ways to price an AI agent:
| Model | Who carries the risk | How it feels to the customer |
|---|---|---|
| Flat per task ($0.25 checkpoint) | The vendor, on big tasks | Predictable, but unfair on small tasks |
| Effort-based (time and compute) | The customer | Fair in theory, scary in practice |
| Per outcome (pay when it works) | The vendor | Easiest to say yes to |
Effort pricing is the one where the customer carries all the risk. That's why it needs the most guardrails. Our guide to AI agent pricing compares these in more depth, and outcome-based pricing covers the third option.
Compare ways to price an AI agent
Model per-seat, per-task and per-outcome pricing side by side before you pick one.
How Did Replit Recover?
It didn't abandon effort pricing. It made it easier to live with:
- Cheaper agent modes that trade capability for cost.
- Model routing (from August 2026), so simple jobs go to cheaper models.
- Big cloud price cuts in August 2026.
- Free Mode inside paid plans, which resets every five hours.
- Lower entry price: Core fell from $25 to $20 a month.
And the business kept growing. In March 2026 Replit said it had more than 50 million users, had tripled its valuation in six months, and was "on track to hit $1 billion in run-rate revenue by the end of 2026."
So the product had real product/market fit. People put up with a confusing bill because the thing worked. That's a lucky position. Most businesses don't get that much forgiveness.
What You Can Steal From Replit
If you sell any kind of variable work, from renovations to legal advice to IT support, here's the playbook.
- Quote before you work. Give a range ("this usually costs $40 to $70") and get approval for anything above it.
- Set spending caps by default. Not as an option buried in settings. On, from day one.
- Don't bill for your own mistakes. If you redo work because you got it wrong, credit it back automatically.
- Offer a cheaper "economy" option. Replit's modes let customers choose speed and power against cost.
- Package the common jobs. Most requests are routine. Price those at a fixed rate and keep effort pricing for the genuinely unpredictable ones.
If you're designing a credit system for AI work, the Credit Pack Designer and the AI gross margin calculator will show you where your margin breaks.
Would I Use Effort-Based Pricing?
Only with a quote first and a cap always on.
Effort pricing is honest about cost. But customers don't buy cost. They buy results, and they want to know the price before they say yes. Get those two things right and effort pricing can work. Get them wrong and even a great product will spend a year apologizing.
See how a rival handled the same pressure in Cursor's pricing story, or compare current plans in the Replit pricing teardown.
Frequently asked questions
What is a checkpoint in Replit?
A checkpoint is a saved unit of completed work by Replit's AI Agent, which you can roll back to. Before June 2025, every checkpoint cost a flat $0.25. Under effort-based pricing, each Agent request creates at most one checkpoint, priced on the time and compute the work took. Simple edits can cost from about $0.06; complex work can cost several dollars.
How much does Replit cost in 2026?
As of October 2026, Replit Core is $20 a month ($18 billed annually) and includes $20 toward more powerful models plus a Free Mode allowance. Pro starts at $100 a month with matching credits, parallel agents and more collaborators. Enterprise is custom. Hosting, compute and storage are metered separately. Check replit.com/pricing for current details.
Why did some Replit users get big bills?
Mostly because the more autonomous Agent 3, launched in September 2025, did more work per request. The Register reported one user who spent $1,000 in a week against $180 to $200 a month before. Users also complained they couldn't see costs until after a task ran, and were charged when the agent looped or failed. Replit offered goodwill credits.
Is effort-based pricing the same as outcome-based pricing?
No. Effort-based pricing charges for the work the AI does, whether or not it succeeds. Outcome-based pricing charges only when a defined result is achieved, such as Fin's $0.99 per resolved support conversation. Effort pricing protects the vendor's margin. Outcome pricing shifts risk to the vendor. Customers usually find outcome pricing easier to accept.
Should I charge by the hour or by effort for my own services?
Charging by effort is really hourly billing in disguise, and it has the same weakness: the customer carries all the risk of the job taking longer. If you do, give an estimate first and agree a ceiling. Better still, package common jobs at a fixed price and keep effort-based billing for genuinely unpredictable work.
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
- Replit, Introducing Effort-Based Pricing for Replit Agent (18 Jun 2025, updated 2 Jul 2025)
- Replit, Effort-Based Pricing Recap (2025)
- Carly, Replit Agent Pricing Explained: Effort-Based Costs and the Backlash (secondary summary)
- The Register, Replit Agent 3 pricing complaints (18 Sep 2025)
- Replit, The Future is Actually Very Human ($400M at $9B, 11 Mar 2026)
- Pricing Innovation, Replit's pricing journey (19 Oct 2026)
- Replit, pricing page (checked 27 Oct 2026)
Facts and prices checked 27 October 2026. Spotted something out of date? Tell us.