Every AI cost overrun I hear about traces back to the same root cause, and it is not the model price. It is who gets to pick the model. Right now the individual contributor picks, and the individual contributor has zero financial incentive to pick anything but the best.
Think about the logic from the dev's seat. "Until they tell me I can't use it, why wouldn't I use the best one available for everything I do?" That is not laziness. That is a rational response to an incentive structure where you want the output but you do not pay the meter. I watched the exact same dynamic play out years ago with databases, where teams ran everything on Aurora at three times the cost of plain RDS because devs just picked the premium option. Nobody could stop it, because the person choosing the resource never saw the bill.
What is masking the severity today is the subscription arbitrage. A capped coding subscription at a couple hundred dollars a month can represent ten thousand dollars of equivalent API spend. These flat-fee accounts are an anomaly from the dawn of LLMs, essentially loss-leader pricing, and they are not going to survive contact with public markets. Investors do not want you losing thousands of dollars per account. The moment that flips to metered billing, the true cost of "always use the best" becomes visible, and the misalignment gets ugly fast. The IC says "I want the output but I can't pay the meter," and there is nobody mediating that tension.
The fix is structural, not a memo asking people to be frugal. You have to move model selection above the contributor and make the cost-quality tradeoff a system decision rather than a personal preference. That means measuring quality per task type so you can confidently route a medium-complexity job to a model at a tenth the price, and it means giving finance leaders the visibility engineering never volunteers. This is the same reason agents are software, not prompts and the same reason the operationalization gap is real work, not a config toggle.
If your AI budget is climbing and nobody can explain why, stop hunting for the expensive model. Look at who is allowed to choose it, and whether they have any reason to choose differently.
Key takeaways
- AI spend explodes because the person choosing the model has no financial incentive to choose a cheaper one.
- The subscription arbitrage hiding real token costs is temporary and will collapse into metered billing.
- Cost governance has to live above the individual contributor or it never happens.
FAQ
Why do developers always pick the most expensive model?
Because they optimize for output quality and bear none of the cost. Until someone tells them to stop, using the best available model for every task is the rational individual choice even when it is wrong for the company.
What changes when subscriptions become metered?
The arbitrage that hides true cost disappears. A flat-fee account that quietly represents thousands of dollars in API-equivalent spend becomes a real line item, and model selection stops being the IC's free choice.
Sources
Related Essays
Agents Give the Signal Humans Are Too Polite to Send
The best prioritization signal in your company dies of politeness. Agents will plus-one the duplicate ticket a human never would, and that changes what engineering sees.
The Coordination Crisis AI Tooling Created
When everyone can build, nobody knows what has already been built. Five teams independently spin up the same Slack bot, and the duplication goes undetected for months.
Users Should Iterate on Agents, Not Developers
Agents are always slightly wrong when first built. The people who know what is wrong are not developers — they are the users who interact with the agent every day.
Drowning in pull requests that need your review? Try Tembo Review, a beautiful AI-assisted PR review tool unlike anything you’ve used.