Who Approved This Token Spend?
Day 3 of 5
Most engineers running agents don’t have a number. Not a budget. Not a ceiling. Not a line item anyone signed off on. It got added to the developer’s tooling allowance, the same bucket as a GitHub seat and a JetBrains license, because that’s the closest category that existed.
That category was designed for tools that cost hundreds of dollars a year. Not $750,000.
Enterprise IT budgets have real limits per slot. Learning and development: $5,000 per developer per year. Hardware: a MacBook amortized over three years. Tooling: $250 for an IDE someone will ask you to justify against a free alternative. When a frontier model agent lands in that last bucket, the math doesn’t work and nobody has officially noticed yet.
It works until the quarterly review. Then someone asks what the tooling line is covering, and the EM is in a room without a good answer. What did we build with this? Who approved it? What breaks if we turn it off?
The pricing question underneath that is harder.
The providers are not charging sustainable prices right now. They’re subsidized by venture capital, propped up by hyperscaler cloud partnerships, and priced to win developer loyalty before the market consolidates. What you’re paying today is not what it costs to run these models. It’s what the providers are willing to accept while they’re still fighting for market share.
Look at what they’re selling. Coding agents. Data analysis. Customer service. Document processing. Operations automation. Every enterprise workflow is a priority market simultaneously. That’s not product focus. That’s a land grab by companies that haven’t yet decided which use cases can support sustainable margins. When that calculus settles, the use cases with clear, measurable ROI survive the repricing. The ones without that story don’t.
The standard rebuttal to all of this is that token costs are collapsing. Prices are dropping faster than Moore’s Law. Today’s math won’t apply next year.
That argument has a problem. The price drops we’ve seen are real, but they’ve come almost entirely on older, smaller models. Frontier model pricing, the models powering the agentic coding sessions we priced out on Monday, has not collapsed. It has shifted modestly while compute demand has exploded. And the efficiency gains from cheaper models get consumed almost immediately by the reality of what enterprise codebases actually look like.
A greenfield mobile app vibe-coded over a weekend fits comfortably in a small context window. A banking platform with twenty years of accumulated business logic, compliance layers, and tribal knowledge baked into fifteen million lines of code does not. Enterprise agents aren’t reading a clean repo with good documentation. They’re ingesting decades of decisions, workarounds, and debt every single session. The context window requirements for that kind of work don’t shrink as models get cheaper. They grow as you ask the agent to do more with the codebase. You’re not running fewer tokens. You’re running more of them on slightly cheaper infrastructure and calling it progress.
Cheaper tokens don’t fix a broken ROI story. They just make it cheaper to produce the same volume of untracked, unauditable output.
MIT studied $30-40 billion in enterprise GenAI investment and found 95 percent of organizations got no return. Some of that is the quality problem we covered yesterday. A lot of it is simpler: the spend was never tied to a measurable outcome, so when the pricing changes and the audit comes, there’s nothing to point to.
The teams that make it through won’t be the ones that spend the most on tokens. They’ll be the ones that can open a log and show what each dollar built.
Tomorrow: the industry has been exactly here before.
Sources
MIT GenAI Divide: State of AI in Business 2025, via The Outpost
Goldman Sachs Chief Economist Jan Hatzius, Atlantic Council, via Gizmodo (February 2026)
NBER Executive Survey (February 2026), via Fortune
Be Atomic covers the infrastructure, primitives, and ideas behind agent-native software development, written by the team building it. atomic.dev