AI Subscription Prices Are Subsidised. Here's What Happens When They're Not.

The $20 you pay for Claude Pro is not what it costs to run Claude Pro. An analysis that hit the Hacker News front page on May 17 (420 points) calculated that a typical Claude Pro subscriber consumes $200–400 of compute per month. Anthropic burns roughly $8 in compute for every $1 of revenue it collects. OpenAI's S-1 filing (May 22) confirmed the same picture from a different angle: the company loses $1.22 for every dollar it earns, projecting a $14 billion operating loss for 2026. And on June 1, GitHub Copilot moves from flat-rate to usage-based billing — after finding that heavy users exhausted five-hour rate limits in 90 minutes.

The AI tools you use every day are priced to acquire customers, not to break even. That works while labs are in a land-grab phase. It does not work indefinitely.

What the numbers mean

Three labs simultaneously burning through investor capital to hold prices below cost is not a stable configuration. Simon Willison noted on May 19 that all three labs are "probing the price tolerance of API customers" — testing the ceiling quietly rather than announcing it.

The most likely outcome is not a single sharp price increase. It is a gradual shift: more capable features behind higher tiers, usage caps tightening on flat-rate plans, and free features migrating to paid ones. GitHub Copilot's move to usage-based billing on June 1 is the clearest signal of the direction. Organisations that have built workflows around a flat $19/month per-seat assumption are about to learn what those workflows actually cost.

What to do differently on Monday

The practical move is not to panic about future prices — it is to audit your current usage against a simple question: which of these tools would I keep if the price were 3x what it is today?

When you answer that honestly, you usually find a clear split: the tasks that save real time and produce genuinely better output, the tasks that are convenient but you could do in five minutes without AI, and the tasks where you use AI simply because it's there and costs nothing extra to prompt. Those three buckets have very different price tolerances.

Knowing which category each workflow falls into before prices change means you make that decision deliberately rather than having it made for you by a bill.

Marcus: finding the high-value core

Marcus is a marketing director at a 30-person SaaS. He uses Claude Pro every day — email drafts, competitive briefs, copy variations for campaigns, prep for difficult conversations with the board. At $20/month, he treats it like a utility.

When the HN analysis came out, he did something simple: he looked at his last 30 days of usage and sorted tasks by how much time they saved versus how replaceable they were.

The results were clearer than he expected. Campaign briefs and the competitive analysis took 30–40 minutes to draft by hand and were genuinely better with AI. Board prep saved him from a particularly bad habit of over-explaining. Those three uses he'd pay $80/month for without a second thought.

Email drafts were a different story. He was using AI to draft routine replies he'd already written a hundred times. At $80/month, he'd stop. He built three email templates instead — 20 minutes of work — and stopped treating Claude like a shortcut for things he already knew how to do quickly.

That exercise left him with a tighter workflow and a clear answer to the pricing question before the pricing question was actually asked.

Elena: recalculating the Copilot budget

Elena runs operations at a 15-person professional services firm. She introduced GitHub Copilot to two developers on her team six months ago. The firm pays a flat $19 per seat per month — she had assumed that was the end of it.

June 1, that changes. Her developers are in the top quartile of usage — the ones who hit rate limits. Under usage-based billing, her actual cost depends on how much they use, not just that they have access.

She doesn't know yet what the new number will be. What she does know is that she approved a tool without modelling usage costs, and she needs to fix that before the first usage-based invoice arrives.

Her plan: ask each developer to log their ten highest-value Copilot uses from the last month, and separately log the habitual low-value ones. Then she'll have a real usage picture — not an assumption — before she decides whether to keep both seats, change who has access, or push for a different billing tier.

The actual planning move

The AI tools market is in a period that has historically preceded a pricing correction: high adoption on subsidised prices, heavy usage concentration in a small number of hands, and labs that have openly told investors they're losing money to grow. You don't need to predict when it changes to prepare for it.

Budget for 2–3x current AI prices in your planning assumptions for next year. Audit your highest-usage tools now and separate the workflows that drive real value from the ones that are just convenient. The professionals who do that now will make better decisions under pressure than the ones who haven't.