The token bill comes due

For a year the only conversation was "AI will replace all developers" — every org hopping onto the bandwagon, letting people go, confident that the math would work out.

Now the math is arriving. And the question isn't about capability anymore. It's about cost.

Boris Cherny burned 7.7B tokens in March and April alone. Mostly Opus 4.6. That's $150,000+ in two months — for one person's usage.

Think about that. A senior engineer's salary, incinerated as tokens.

The economics only work if AI replaces multiple humans. But what if it doesn't? What if it just makes each human 2x more productive while adding a $75k/year token habit? You've now got an engineer who costs $250k instead of $150k.

Here's the thing nobody mentions: not all automations actually save money. Some tasks, a human doing it once a day is still cheaper than the tokens required to AI-do it every time. Token optimization is about to become a real skill — using AI strategically rather than reflexively.

But the bigger risk is the one nobody talks about: these prices are subsidized. VCs are footing the difference between what it costs to run these models and what you're being charged. Growth mode. Capture market share. Make the unit economics look palatable until they don't have to.

Right now, Anthropic is incentivized to keep prices low. But once the dependency is baked in? Once entire workflows assume Claude at a certain price point? The bill comes due. Companies spending $150k per power user will hit a ceiling and discover that the thing they thought was replacing developers is actually just a very expensive line item that keeps getting more expensive.

The second-order effect: all this AI tooling is being built on the assumption that current API prices are the new normal. They're not. They're the intro rate.

---

Maybe the real question isn't whether AI is cheaper than humans. It's whether AI-dependent companies can survive their own token bills.