OpenAI killed its fixed-tier pricing for Codex and replaced it with pay-as-you-go. In the same week, Coder raised $90 million in Series C funding from KKR. And seven of the nine text models released in March 2026 were open-weight, free for anyone to run. Three data points. One conclusion: the $20-per-seat subscription that built the SaaS era is under siege, and the attackers are coming from every direction at once.
Here is the thing nobody wants to say out loud. Most SaaS companies set their price once, slapped it on a per-seat grid, and never thought about unit economics again. That worked when software was static and usage was predictable. AI coding tools are neither. Another developer on the same team might use 200 tokens a week. Charging both $20 a month is either a steal or a rip-off. It is never the right price for both.
I think this is the most important pricing shift in developer tools since Twilio proved that APIs could be billed like electricity. But the data is mixed on whether pay-as-you-go actually replaces seats or just layers on top of them. The truth, as usual, is messier than the headline.
The Meter Principle
Here is the framework. I call it The Meter Principle: when the cost to serve a customer varies 10x or more between your lightest and heaviest user, flat pricing becomes a subsidy from the light users to the heavy ones. Eventually the light users leave, the heavy users stay, and your margins collapse. The meter fixes this by tying revenue to consumption.
Think of it like your electric bill. Nobody pays a flat monthly fee for electricity regardless of whether they run one lamp or a server farm. The meter exists because usage variance is enormous. AI coding tools have the same variance problem. That is a 20x spread in the same product category. The variance is the signal.
The Meter Principle has three rules. First, if your heaviest user costs 10x more to serve than your lightest, you need a meter. Second, the meter should be legible. Tokens, requests, or credits, pick one unit and make it obvious. Third, never surprise the customer with the bill. Usage-based pricing only works when the customer can predict their spend within 20% before the invoice arrives.
Dan Martell would call this the Replacement Ladder for pricing models. You do not rip out seats overnight. You start by metering overages (GitHub Copilot Pro+ charges $0.04 per extra request). Then you shift the default to usage-based for new plans. Then you sunset the old tiers. OpenAI just jumped to step two with Codex. Everyone else is still on step one.
The Hard Way vs. The Easy Way to Model This
Let me show you exactly how to think about this if you are a founder building or selling AI tools right now.
The hard way: you set a flat price of $20 per seat per month because that is what Cursor, Windsurf, Claude Code Pro, and Augment Indie all charge. According to getdx.com's April 2026 analysis, a 500-developer team on GitHub Copilot Business pays $114,000 a year. The same team on Cursor Business pays $192,000. You pick a number, multiply by headcount, and hope your margins hold. The problem is that your infrastructure costs are not flat. They scale with tokens consumed. Every power user who maxes out their allocation is eating your gross margin for breakfast.
The easy way: you meter usage and let revenue track costs automatically. Claude Code's API charges $3 per million input tokens and $15 per million output tokens on Sonnet 4.6. Opus 4.6 runs $5 and $25 respectively. Your revenue goes up when your costs go up. Your revenue goes down when your costs go down. You are literally just matching the shape of your expense curve to your income curve. Stupid easy, once you see it.
But here is where most people get it wrong. They think pay-as-you-go means unpredictable revenue. It does not have to. The move that is insane right now is the hybrid model. You charge a base seat fee for access, then meter heavy usage on top. GitHub Copilot Pro+ does exactly this: $39 per month gets you 1,500 premium requests, then you pay per request after that. Augment Code charges $60 per user per month for Standard, then auto-tops-up at $15 per 24,000 credits beyond the cap.
Let me walk you through the math on a napkin. Say you have 50 developers. Now imagine 10 of those developers are power users burning 5x the average tokens. The other 40 barely touch the tool. The AI Corner reported in March 2026 that the average user actively uses only 42% of their paid AI subscriptions. You are paying full price for 50 seats but getting real value from maybe 20.
Under a metered model, those 40 light users might cost you $5 to $10 each per month. The 10 power users might cost $100 to $200 each. Your total spend could drop from $240,000 to somewhere around $120,000 to $150,000. Or it could go higher if everyone becomes a power user. Either way, the price reflects reality instead of a guess.
Now here is the part that should make every SaaS founder nervous. Seven of nine text models released in March 2026 were open-weight. That means any developer can run them locally or through a cheap API provider. If you are charging a flat $20 per month for access to a model that someone else is giving away, your pricing power evaporates overnight. The meter is not just better economics. It is a survival strategy. When the model layer commoditizes, the only defensible pricing is on the value delivered per unit of work, not on access to the model itself.
It is unclear whether pure pay-as-you-go will dominate or whether the hybrid model wins long term. Security experts at Endor Labs have pointed out that AI-generated code contains vulnerabilities 62% of the time by default, which means companies need to budget for review tooling like CodeRabbit at $12 per user per month on top of the coding assistant. IP lawyers including Jeffrey Gluck have flagged that AI outputs may contain copyrighted training data, creating legal costs that do not show up on any pricing page. The real unit economics of AI coding tools include a lot more than the subscription fee.
My read on this: the winners will be the companies that meter transparently, bundle review and security into the price, and make the bill predictable. The losers will be the ones clinging to flat per-seat pricing while their margins get eaten alive by power users and their light users defect to free tiers. GitHub Copilot Free already offers 2,000 completions and 50 chat requests per month at zero cost. Bolt.new gives away 1 million tokens per month. The floor is free. You cannot compete with free on a flat fee. You compete with free by charging for value above the floor.
2031
Pull back five years from now. Here is where this gets interesting at the structural level.
The SaaS industry spent 15 years optimizing for one metric: Annual Recurring Revenue. ARR is the number that VCs fund, that boards track, that founders tattoo on their foreheads. Seat-based pricing was perfect for ARR because it was predictable, stackable, and easy to model. Usage-based pricing breaks that entire mental model.
Coder raising $90 million from KKR in the same week OpenAI went pay-as-you-go is not a coincidence. It is a signal that enterprise buyers want AI dev tooling but want it priced like infrastructure, not like software licenses. KKR is a financial engineering firm. They understand metered revenue. They understand that consumption-based models can actually grow faster than seat-based models because there is no ceiling on per-customer spend.
The asymmetric advantage belongs to founders who build their financial models around gross margin per token rather than ARR per seat. Nvidia nearly went bankrupt in the late 1990s before it figured out that its real business was not graphics cards but compute cycles. The same reframing is happening in developer tools right now. The product is not the seat. The product is the unit of work completed.
Five years from now, I expect the $20 per seat tier to still exist, but only as an on-ramp. The real revenue will come from metered usage above the base. Companies like Augment Code, which already cap plans at 20 users and charge overages, are building for this future. Companies that treat every seat as equal revenue are building for a past that is already ending.
The compounding effect matters here. As models get cheaper (open-weight releases are accelerating this), the cost to serve each token drops. But the number of tokens consumed per developer goes up as AI agents handle more complex, multi-file tasks. Revenue per developer can grow even as cost per token shrinks. That is a flywheel. Seat-based pricing cannot capture it.
What to Build This Weekend
Stop theorizing. Here is what to do in the next 48 hours.
First, audit your own AI tool spending. List every coding tool you pay for. Write down the monthly cost and your actual usage. If you are using less than 50% of any tool's capacity, you are overpaying. The AI Corner's data says the average is 42% utilization. You are probably in that bucket.
Second, test a metered alternative. If you are on Cursor Pro at $20 per month, try Claude Code's API pricing for one week. Track your token usage. Multiply input tokens by $3 per million and output tokens by $15 per million for Sonnet 4.6. Compare that number to your flat subscription. You might be shocked at how much cheaper (or more expensive) your actual usage is.
Third, if you are building a product, model your pricing both ways. Use Fabricate v2.0 to spin up a quick prototype of a usage dashboard. It turns a single text prompt into a deployed web app. Build a calculator that shows customers their estimated monthly cost under seat-based versus metered pricing. Make the meter visible. Transparency is your competitive advantage.
Fourth, stress-test your API workflows. Use Luzo, the open-source visual debugger, to map out every API call in your AI coding pipeline. Find where tokens are being wasted. An ounce of optimization in your prompts is worth a pound of savings on your bill.
Fifth, read the fine print on your current tools. GitHub Copilot Enterprise requires GitHub Enterprise Cloud at an additional $21 per user per month. Augment Code caps at 20 users per plan. Cursor does not pool allocations across team members. These hidden costs change the math completely.
The Meter Principle is not complicated. Match your revenue shape to your cost shape. Charge for what is consumed. Make the bill legible. The founders who internalize this in 2026 will own the developer tools market by 2031. The ones who do not will be subsidizing power users until the money runs out.
Get your reps in. Build the calculator. Run the numbers. The pricing model you choose today determines the company you become tomorrow.