There is a spreadsheet making the rounds in boardrooms right now. On one line: the fully-loaded cost of a senior software engineer. On the other: what it costs to run a coding model that writes the same kind of code. The gap looks enormous, and it is driving real decisions about real people. The problem is that the spreadsheet is measuring the wrong two things.

Let me run the actual numbers, because the headline comparison is seductive and mostly wrong.

The number on the left

A senior software engineer in the US has a median total compensation around $250K, base plus bonus and equity, and at top-tier firms it clears $300K. Then add what the employee never sees: employer payroll taxes, benefits, equipment, the recruiter who found them. The fully-loaded cost the company actually carries runs higher still. Call it $250K for a conservative working number.

The number on the right

Now the tokens. Take an engineer working a coding agent hard, all day. Agentic coding is input-heavy — the model reloads context, reasons, retries, and the codebase gets fed in over and over. A heavy day might burn 10 million tokens, split maybe 80% input and 20% output.

At Claude Sonnet 4.6 rates ($3 per million input, $15 per million output), that day costs about $54. Across 250 working days, roughly $13,500 a year. Run everything on the more expensive Opus 4.8 ($5 in, $25 out) and it's about $90 a day, or $22,500 a year. Prompt caching pulls the real figure lower. Even if you double the usage, you are under $50K.

So there it is. A $250K engineer versus $20K of tokens. Ten times cheaper, maybe more. If that is the comparison, the conclusion writes itself, and a lot of companies have already written it.

Why the comparison is wrong

The two numbers are not the same kind of thing. The $250K buys an engineer. The $20K buys coding output — and coding output is not an engineer.

Start with what an engineer actually does. Writing code is maybe 30% to 50% of the job. The rest is figuring out what to build, deciding what not to build, untangling an ambiguous requirement, coordinating across teams, reviewing other people's work, and debugging the production incident at 2 a.m. that has no clean reproduction. The model is good at the first part and cannot do most of the second.

Then look at the part it can do. On the coding itself, a good agent handles somewhere between 30% and 70% of the work, depending on how novel and well-specified the task is. Multiply that through: the model covers roughly 15% to 35% of the actual role. Not the role. A slice of it.

And that slice is not free, even after you pay for the tokens. AI-written code still needs a senior human to read it, catch the plausible-but-wrong logic, and own the result when it ships. Reviewing machine output at volume is real work — call it 0.3 to 0.5 of a senior engineer's time, which is $75K to $125K of human cost you just put back on the table. Add the rework when something slips through, the incidents, and the ongoing cost of maintaining the prompts and evaluations that keep the agent useful.

Rebuild the spreadsheet honestly and the picture inverts. You cannot buy back a whole engineer for $20K. You can shave 15% to 35% off the effort of the role, net of a meaningful review tax. That is a real productivity gain. It is not a deleted headcount. The cheap-token number answered a question nobody should have asked: not "what does an engineer cost" but "what does it cost to generate code that still needs an engineer behind it."

Meta is the tell

If the knowledge were already captured and cheap, you would not need to pay people to hand it over. Meta is doing exactly that, in public.

Through 2026, Meta began recording employees' screens and keystrokes across tools like GitHub and Slack to teach AI agents how to do end-to-end work. It reassigned thousands of engineers to manufacture training data — writing coding problems, grading machine-written code. And it did this in the same window it cut roughly 8,000 jobs while posting record profit.

Read those three facts together. A company convinced that AI already replaces engineers does not spend its engineers' time teaching the AI what they know. The training program is an admission: the expensive part of the job lives in people's heads, it has not been captured yet, and capturing it costs real engineering hours. The "replacement" is being hand-built by the people it is meant to replace.

The bill the company books

Inside one company's ledger, the long-term costs are already visible. Cut the engineers and you lose the context that makes the tools useful in the first place — the people who knew why the system was built the way it was. Stop hiring juniors because the agent does junior work, and you break the pipeline that produces seniors; in five years there is no one left who can review the machine. Klarna learned the near-term version the hard way: after cutting customer service deep on the promise of AI, it started rebuilding the human side, because the savings came with a quality cost it had not priced.

The bill the company does not book

The larger cost lands outside the spreadsheet entirely, and it is worth being plain about.

Putting people out of work at scale, on purpose, to book a margin, is a choice with a moral weight that "efficiency" language is designed to hide. These are livelihoods, not line items. A firm gets to call it a headcount reduction; the person calls it a mortgage they can no longer pay. Treating that as a rounding error because the spreadsheet looks clean is the part I cannot get past.

It is also self-defeating in a way that should bother even the people who do not care about the first part. You cannot sell software to an economy you have helped defund. Tech workers are also tech customers, taxpayers, and the demand side of the market every one of these companies sells into. Strip enough spending power out of the middle of the economy and the gains concentrate at the top while the base that buys the product erodes. Henry Ford understood the cruder version of this a century ago: pay people enough to buy the cars.

The honest counterargument is that automation has always displaced workers and produced new jobs on the other side, and over a long enough horizon it has. The difference this time is the speed, the breadth across white-collar work at once, and who captures the gains. When the productivity surplus flows almost entirely to capital and the transition is measured in quarters instead of decades, "it worked out before" is a bet, not a guarantee.

The choice nobody is naming

There are two ways to hold an AI that does 15% to 35% of an engineer's job. One: keep the team, and let each person ship more — the experiments and half-built products that never made it off the backlog because there was never time. That is what investing in productivity actually looks like. Two: keep the output flat, cut the people, and book the difference.

The first treats an immature technology as a multiplier for the people you already have. The second treats it as a reason to push risk onto your workforce and the economy around you, for a near-term number the all-in math already says is smaller than it looks.

The spreadsheet says option two is free money. Ask the thousands of Meta engineers grading their replacements' code whether it feels free.