Modern Pricing & UBB
February 27, 2026 pricing

The $20/Month AI Subscription Was Always a Lie

Let's talk about the greatest fiction in modern software pricing: the flat-rate AI subscription. Twenty bucks a month. Unlimited intelligence. All-you-can-eat inference. The Netflix model, but for the actual robots taking your job.

It was never going to work. Everyone who could do basic math knew it. And now — slowly, painfully, with the grace of a shopping cart rolling into a parking lot — it's falling apart in public.

The Math Was Broken From Day One

Here's the problem with selling "unlimited AI" at a flat rate: the cost to serve isn't flat. Not even close. A power user running agentic coding workflows — the kind where the AI reads files, writes code, hits an error, fixes it, re-runs tests, loops back — can burn 100,000+ tokens on a single complex task. At standard API rates for a frontier model, that's $1.50 to $3.00 per task. One task. Do that a few dozen times a day and your $20/month customer is costing you $500.

That's not a rounding error. That's a business model error.

The whole thing was propped up by two things: venture capital subsidies and the assumption that most users would be light users. Both assumptions are now visibly wrong. AI-native app spend jumped 108% in a single year according to Zylo's 2026 SaaS Management Index — large enterprises surged 393%. Usage is exploding. Heavy users are multiplying. The math gets worse every month.

The Body Count Is Starting to Pile Up

This isn't theoretical anymore. Companies have tried flat-rate AI and learned its limits the hard way — in public, on the record, with receipts.

Windsurf is the most dramatic case. Windsurf ultimately had to shut down and sell off assets after burn rates outstripped growth. It got squeezed by rising compute costs on one side and aggressive flat-rate competition on the other. Classic race to the bottom, except the bottom was bankruptcy.

Claude Code's "Max" unlimited tier — priced at $200/month — was one of the most ambitious experiments in unlimited AI pricing. Bold price point, multiple model tiers, even offloaded some processing to users' own machines. Still didn't work. Some users were racking up $25,000+ per month in compute value on a $200/month subscription. The unlimited tier got rolled back. Obviously.

Cursor posted "important pricing updates." Translation: we discovered our best customers were using us like a public utility and we can't sustain it. They're not alone — this is now a genre of announcement.

The Prisoner's Dilemma That's Eating Everyone

Here's what makes this genuinely hard: it's not stupidity. It's game theory. If you switch to usage-based pricing and your VC-subsidized competitor keeps offering unlimited, you lose customers. So everyone keeps the flat-rate lie going a little longer, hoping the other guy blinks first.

Nobody blinks. Everyone bleeds.

Meanwhile, 92% of SaaS companies have launched or plan to launch AI features. They're all going to discover this same problem. The token costs of newer "reasoning" models — the ones doing multi-step, chain-of-thought work that users actually find valuable — don't drop as fast as the benchmarks imply. Because users migrate instantly to the best available model, and the best model is never cheap.

You get a short squeeze: consumers expect Netflix pricing, but the underlying compute economics look more like cloud infrastructure. AWS doesn't offer unlimited EC2 for $20/month. Neither can you.

So What Actually Works

The exits from this trap are well-understood, even if nobody likes them:

The conclusion is unavoidable: usage-based billing isn't a niche preference for infrastructure nerds. It's the only sustainable model for AI products at scale. Every company that's tried to hide this reality behind a flat monthly fee has either burned through its runway or is currently doing so.

The $20/month AI subscription was always a lie. The market is finally starting to admit it.

Sources

← Pricing Strategy · All posts