9.4x. That's how much more code churn - rewrites, reversions, and deletions - regular AI coding tool users are producing compared to colleagues who don't use AI, according to a January report from GitClear. GitClear also found that churn exceeded actual productivity gains by more than double. That ratio is what happens when you optimize for output volume instead of output quality.
The practice driving it is called "tokenmaxxing." Tokens are the small units AI models process - roughly three-quarters of a word each, about 750 per page of text. When developers tokenmaxx, they prompt tools like Cursor or Claude Code to generate as much code as possible in one shot, treating raw volume as a proxy for productivity. A TechCrunch analysis compiling data from several engineering analytics firms found the approach is backfiring across the industry.
The Acceptance Rate Trap
Waydev, which tracks engineering metrics for companies with 10,000+ engineers, measured something telling: AI-generated code carries an 80-90% initial acceptance rate. Developers merge it, it looks fine, it ships. But four to six weeks later, that acceptance rate collapses to 10-30% as problems surface and teams start rewriting. The cost doesn't disappear at merge time - it just moves into the future.
Faros AI put harder numbers on this in a March 2026 report covering two years of customer data: code churn increased 861% under high AI adoption. That's not noise. That's a sign that a significant share of AI-generated code being merged isn't actually staying in production - it's just creating a cleanup backlog that doesn't show up in PR count metrics.
2x Throughput, 10x Token Cost
Jellyfish tracked 7,548 engineers through Q1 2026 and found that engineers with the largest token budgets did produce more - roughly 2x the pull request throughput of lower-budget peers. But they did it at 10x the token cost. Running frontier coding models at maximum output isn't free, and the efficiency ratio is ugly: you're paying ten times as much for twice the output, while also accumulating hidden rewrite debt.
The pattern matters most for teams paying per-token API costs on tools like Codex or Claude. A team that looks productive by PR metrics may be spending a large multiple of what a more targeted approach would cost - and banking technical debt in the process.
Who Gets Hit Hardest
The Waydev data points to junior engineers as the most exposed group. They accept more AI-generated code than senior developers and face greater consequences when that code needs rewriting. Senior engineers tend to prompt more precisely, catch bad output faster, and discard more aggressively. They use AI tools to accelerate work they already understand how to do.
Junior developers - the ones these tools were supposed to help most - are more likely to merge code they don't fully understand. When that code surfaces bugs later, the person debugging it may not have enough context to fix it cleanly, making the rewrite cost compound.
Atlassian's $1 billion acquisition of engineering intelligence startup DX is a signal that the industry recognizes it has a measurement problem. Lines of code and pull request counts are not telling the real story, and companies are paying heavily to figure out what AI-assisted productivity actually means.
The developers consistently getting good results from these tools are treating them as precise instruments, not firehoses: small, targeted prompts, incremental tasks, careful review before merging. The mental model isn't "generate everything and cut later" - it's "write less, ship more of what you write."