Waydev: The Token Race Is Cutting Returns on AI Coding and Inflating Rework

The new cult in teams that lean heavily on AI coding is not release quality but the volume of tokens burned. The problem is that more code and more PRs do not necessarily mean more useful work: a significant share of that volume later goes into rewriting.

How tokenmaxxing works

The term tokenmaxxing describes a simple logic: if a developer spends a lot of tokens in Claude Code, Cursor or Codex, they are supposedly working more productively. Inside companies, large AI compute allowances have become a status symbol, much as lines of code or commit counts once were. But it is the same old mistake in new packaging. Tokens are an input to the process, not an outcome. They show how many resources were burned, but they do not answer whether the team now has more reliable, useful and maintainable code.

According to Waydev, which works with roughly 50 companies and more than 10 thousand engineers, managers often see a very attractive top-level picture: 80–90% of AI-generated code is initially accepted and lands in the repository. But then comes the less visible part of the work: fixes, reverts and rewrites over the following weeks. Once you account for that tail, the real share of code that actually "sticks" drops to 10–30%. That is where the illusion of instant productivity breaks down.

Metrics versus illusions

Several engineering analytics companies are now trying to measure this picture. Demand has grown so much that big players have started investing in such tools too: in 2025, Atlassian bought the startup DX for $1 billion to better measure the return on AI agents in development. The general conclusion across platforms is similar: the volume of code shipped is growing, but the share of durable output is not.

• GitClear reports that regular users of AI tools show code churn 9.4 times higher than developers without AI, although the year-over-year growth, once adjusted, looks much more modest, at around 25%.
• Faros AI, in its March 2026 report, recorded that with a high share of AI in development, PR size grew 51%, bugs per PR grew 28%, median review time grew fivefold and code churn jumped 861%.
• Jellyfish, in a sample of 7,548 engineers, found that the developers with the biggest token budgets produce more PRs, but the habit quickly becomes unprofitable on cost.
• In Jellyfish data, the top 20% by token spending shipped an average of 23 merged PRs per quarter versus 11 for the bottom 20%, but spent about $1,822 on it versus roughly $3.

The price of excess code

The main problem is not that AI writes a lot, but that the excess volume shifts the cost of the work down the pipeline. The team generates drafts faster, but then pays for reading, reviewing, integrating and maintaining that code. If PRs get larger, bugs more frequent and reviews longer, part of the "speedup" is simply deferred to the future as technical debt.

This is especially visible with junior developers: they more often accept AI suggestions without heavy filtering and then more often return to those changes. The situation is better for senior engineers, because they set context more precisely and cut junk at the input more aggressively. Even so, marginal returns decline. Jellyfish puts it almost bluntly: at the top end of the curve, developers spend almost 10 times more tokens to get roughly a twofold increase in throughput. That is no longer a story about efficiency but about very expensive overclocking.

"This is a new era of software development, and companies will have to adapt regardless," says

Waydev CEO Alex Circei.

What this means

AI coding is not going away, but the cult of maximum token spending looks like a poor management metric. If companies want real productivity, they will have to look not at the amount of generated code but at how many changes survive review, do not break production, do not come back for rewriting and pay back their cost in tokens.