AI coding tools are killing themselves over bad pricing

The free tier death cult

One user consumed tens of thousands in model usage on a $200 plan.

That’s from Anthropic’s own admission when they killed Claude Code’s unlimited tier. Not $200 worth of usage. Tens of thousands of dollars on a $200/month plan.

TechCrunch said Cursor hit $500 million ARR. Fastest growing application ever, but still losing money on every power user. When Anthropic rolled out new rate limits, they pointed to their biggest power users as the problem. Some were “running it continuously in the background 24/7”.

These aren’t edge cases. These are the natural consequences of unlimited AI.

This is the paradox of AI economics: the better your product gets, the faster you go bankrupt.

At least if you don’t price it right.

Replit’s reality check

The Information got their hands on Replit’s actual financials. The numbers are magnificent, but not in a good kind of way:

  • February 2025: 36% gross margins
  • April 2025: -14% gross margins

In two months, they went from barely profitable to paying customers to use their product.

They discovered what every power user eventually discovers: if you’re not rate-limited, you might as well run your agents 24/7. The marginal cost to you is zero. The marginal cost to Replit is everything.

The 10B token man

This one’s from Anthropic’s Claude Code experiment, documented by viberank:

One user used 10 billion tokens in one month.

Why? I believe every code commit triggers a complete rewrite. Every rewrite triggers another commit. An infinite loop of AI talking to itself on Anthropic’s infrastructure, which eventually caused Anthropic to kill it off.

The Windsurf warning

TechCrunch reported what everyone in the valley already knew: Windsurf and Cursor are “absolutely bleeding VC money on every call”.

The specifics, according to their sources:

  • Revenue per user: $20/month
  • API costs for power users: $80-200/month
  • Margin on heavy users: -300% to -500%

Windsurf sold for parts. $82 million ARR but nobody wanted it at any price…

When buyers pay billions for your employees but won’t touch your revenue, you’re not running a business but an extremely expensive training program – so what are you selling exactly?

The industry-wide delusion

Nicholas Charriere, founder of Mocha, told TechCrunch the quiet part out loud: “Margins on all of the ‘code gen’ products are either neutral or negative. They’re absolutely abysmal”.

Read that again – he said all, not some. All.

At least Ralph understood he’s in danger…

This is an entire industry subsidizing compute at venture capital’s expense. Like Uber before, and Moviepass before that.

That glorious dumpster fire that let you watch unlimited movies for $9.99/month, and lost $150 per customer per month and thought they’d make it up on volume.

Your path forward

Path 1: The Anthropic retreat

Charge 10x more ($200/month when competitors charge $20).

Add every optimization possible.

Auto-scale from expensive models to cheap ones.

Offload processing to user machines.

Still lose money. Kill unlimited. Pretend it was always the plan. ✌️

Path 2: The Windsurf exit

Burn VC money to subsidize users.

Collect training data.

Hope someone acquires you for the team before the music stops.

Some owners of Windsurf won at that game – the rest lost.

Path 3: The Replit pivot

  • January: Outcome-based pricing
  • April: Lose money on every task
  • July: Switch to effort-based pricing
  • Result: 4-5x price increases overnight

The CEO of replit insists unit economics aren’t a problem. The Information’s numbers suggest otherwise.

If Anthropic couldn’t make it work, why will you be able to?

Every founder tells me their free tier is essential.

“It’s our funnel, it’s how we grow”.

Sure, cool, yeah. But if Claude Code at $200/month with Anthropic’s own models couldn’t make unlimited work, why do you think you can do it at $20?

You know you can’t.

Your investors know you can’t (or they’re cheating themselves).

Everyone pretends because admitting it would mean admitting the entire business model is broken.

Your free tier isn’t a growth engine. It’s a time bomb.

And the clock’s ticking louder than you think.


Posted

in

,

by

Tags:

Comments

5 responses to “The free tier death cult”

  1. Na Avatar
    Na

    What’s the alternative? How do you hope to get users to pay when they haven’t seen any proof with AI output and when a reasonable standard of output hasn’t been established? Or is this leaning more towards dev tools and vibe coders?

    1. Arnon Shimoni Avatar
      Arnon Shimoni

      Mostly geared towards the dev tools. But I also firmly recommend pricing outcomes and not flat fees

    2. S Avatar
      S

      The best alternative, that every other industry in the world uses all the time but for some reason media and tech are very slow to learn, is to sell individual units. SaaS subscriptions create these sorts of distorted markets. SaaS subscriptions are always about subsidizing heavy users at the expense of people who forgot they signed up.

      LLM quality is going down for the same reason that Netflix Original quality has been going down: there’s an not a direct relationship between the user and a discrete product. The user isn’t buying 10 individual movies a month, he’s paying a flat rate and he’s either hoping that he can find 10 movies that will make his 9 bucks worth it, or he’s watching as much as possible to guarantee that at least the quantity is worth it.

      The closer put the price to the smallest unit of goods, the better the price will be. In this case, you price per prompt is good; price per token is better.

  2. Geffen Avatar
    Geffen

    How do you price for outcome when quality (1) matters and (2) is subjective?

    1. Arnon Shimoni Avatar
      Arnon Shimoni

      You need to go deeper with your customers and potentially also have “discriminatory” pricing based on what the customer needs/values

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.