The NBER's study of more than 100,000 developers should end the "10x engineer" fantasy for good: a 180% increase in commits produced a 30% rise in releases and no measurable lift in end-user adoption. Meanwhile, Plandek's Q4 2026 benchmark across 2,500 teams found token spend on coding tools climbing 10–13x while productivity hopes stalled.
Generation got cheap; ownership didn't
AI has collapsed the cost of producing code, not the cost of owning systems. Cursor studies show velocity gains that arrive with "long-term complexity" attached. Anthropic's analysis of 400,000 agent sessions found the same model succeeds roughly twice as often in expert hands as in novice hands. The hard work — requirements, integration, security, review — never moved into the model. It's still sitting on the shoulders of the same senior engineers, now buried under an avalanche of plausible, unverified diffs.
The bottleneck moved; the budget didn't
This is where it's heading, and it's a CFO problem as much as a CTO problem. The unit economics of engineering are flipping from headcount to inference spend plus review capacity. If tokens rise 10x while releases move 30%, the naive "more code = more value" model is already dead. Cheap generation means the constraint becomes judgment: who decides what to build, who verifies what was generated, and who accepts liability when it fails in production. We are not running out of code. We are running out of people qualified to review it.
Rebuild the measurement, not the team
Measure releases and adoption — not commits, not velocity. Re-price every coding tool against delivered change, not seat count. And expect the hiring market to bifurcate: junior translation work compresses, senior verification work gets scarcer and more expensive. The teams that treat code generation as a free good will pay for it as an unmanageable estate.