Skip to content
AI-Native PM

A real incident

OpenAI priced its flagship plan on a guess and lost money on every subscriber; the chapters next door turn cost, speed, and price into decisions you make before that bill arrives.

01 / 09
A real incident

OpenAI set its ChatGPT Pro plan at $200 a month, then admitted it was losing money on every subscriber.

The model was the best on the market and users loved it; the part that broke was the price, set on a guess about how much people would actually use it.

TechCrunch, January 2025, on Sam Altman's public admission

02 / 09
What happened

The CEO who set the price posted the result himself.

POSTED TO X, JAN 2025

insane thing: we are currently losing money on openai pro subscriptions! people use it much more than we expected.

He had picked the price himself.

Pro subscribers leaned on the heaviest reasoning models and Sora far more than the flat fee assumed, so each active user cost more to serve than they paid.

03 / 09
What actually failed

The model worked. The economics around it were never designed.

The price was set before anyone knew what one heavy session actually cost to serve.

Usage carried no ceiling, so the meter ran as far as each user wanted to push it.

A flat monthly fee could not track a bill that moved with every request.

A better model would not have saved the plan, because the gap was never in the model.

04 / 09
The pattern

This is the default failure of AI products, not a one-off.

$200/mo

OpenAI's Pro plan still ran at a loss

-30%

Cursor's gross margin after paying for inference

40%

Anthropic's cut 2025 gross-margin forecast as inference ran over

Across the market, revenue climbs while the cost to serve each user climbs faster, and the margin is what gets squeezed.

Altman on X (2025); Foundamental estimate via Ed Zitron (2025); Anthropic 2025 guidance

05 / 09
The one idea

Margin is a product decision you design before launch, not a number you read off the monthly invoice.

Cost, speed, and price are choices you make while building the feature, so every answer it returns costs less than it is worth.

06 / 09
What this part builds

The part gives you the levers that turn a feature's cost from a surprise into a number you designed.

Cost per task

Add up every call, retry, and judge pass one finished answer triggers.

Routing cascade

Send each request to the cheapest model your evals still accept.

Caching

Serve repeated reads and answers for free, and evict them when facts change.

Price that tracks the meter

Charge in a currency that moves with cost, tested against your heaviest user.

07 / 09
The objection worth answering

"Inference gets cheaper every year, so the margin fixes itself."

Prices per token do fall, but usage per user rises to meet the cheaper rate, and heavy features pile on retries, judge passes, and longer context that eat the savings. The bill follows demand, not the price sheet.

Waiting for hardware to rescue a margin hands your P&L to someone else's roadmap. The decisions that hold cost below value are yours to make now, in routing, caching, and price.

08 / 09
What actually fixes it

Each gap in the story maps to a move you can build.

Price set before cost was knownCost per task

→ Total every call and retry one answer triggers before you set a price.

No ceiling on heavy useRouting cascades

→ Route each request to the cheapest model your evals admit, and watch the escalation rate.

Same work paid for twiceCaching

→ Serve repeated reads and answers from a cache, evicting them when the facts change.

Flat fee, variable billPricing

→ Charge in a currency that tracks the meter, tested against your heaviest user.

Every one of these was decidable before launch, not after the invoice arrived.

09 / 09
From story to part

Every gap in this story is a chapter ahead.

Eight chapters take a feature's economics from a monthly surprise to a designed system, and end in one signed Inference Budget per feature, with alarms that page a name.

Start with chapter one →
↑ The Builder's StackThe Frontier map →