AI tools are multiplying engineering output fast enough to shift where the bottleneck actually lives

Anthropic engineers now ship eight times as much code per quarter as they did compared to the 2021 to 2025 period. That figure comes from Fiona Fung, and it is not a projection or a model output. It is an observed internal metric. Alongside it, Fung notes that most commits at the company are now AI-assisted. The shape of what that implies is structural: coding is no longer the bottleneck.

When output volume rises eightfold, the constraint shifts. The work that slows things down is no longer writing code. It is reading, reviewing, and verifying what the tools produced. That is a different engineering problem than the one most teams are currently organized to solve.

Clay Bavor adds a data point from outside Anthropic that gives the productivity numbers some external texture. Engineers who lean heavily into coding agents self-report productivity gains of between three and twenty times in features shipped. The range is wide, and Bavor does not flatten it. The variation almost certainly reflects how deeply individual engineers have actually integrated these tools into daily work, not just whether they have access to them.

Top engineers who are really leaning in to cla codeex and so on are spending more than $100,000 on a run rate basis on tokens per year. Clay Bavor

The token spend Bavor cites is its own kind of signal. Top engineers who are fully committed to the approach are spending more than $100,000 per year on tokens at a run rate. That is a meaningful personal expenditure, or a meaningful line item for whoever is covering it. It suggests the engineers seeing the largest gains are not dabbling. They are running the tools hard, continuously, as a core part of the job rather than an occasional assist.

Taken together, the Fung and Bavor data points describe the same underlying dynamic from two angles. One is an internal Anthropic measurement of aggregate throughput. The other is a self-reported range from engineers across a broader population. Neither figure is a controlled study, and self-reported productivity gains always carry some inflation risk. But an eightfold increase in shipped code is not the kind of number that disappears under scrutiny.

What the data does not yet resolve is what happens downstream of the throughput gain. If most commits are AI-assisted and volume has multiplied eightfold, the humans reviewing that code are facing a workload that scales with output, not with headcount. Fung’s framing, that coding is no longer the bottleneck, points directly at this. The productivity win at the generation layer may be accumulating pressure at the verification layer faster than teams have noticed. Whether the same tools that accelerated output can also accelerate review is an open question the current evidence does not answer.

The pattern is early and the data points are few. But an eightfold throughput increase at one of the companies building these tools, corroborated by wide-range self-reports from engineers elsewhere who are spending heavily to stay in the workflow, is worth treating as more than an anecdote. The engineers who have committed most fully to the approach are reporting gains that, if they hold under scrutiny, will force a reconsideration of what a software team actually needs to look like.