‘Tokenmaxxing’ is making builders much less productive than they believe

There’s an outdated noticed in control: What you measure issues. And, normally, you get extra of no matter you’re measuring.

Tool engineers have debated productiveness metrics for many years, beginning with strains of code. However as the brand new technology of AI coding brokers delivers extra code than ever, what their managers needs to be measuring is much less transparent.

Huge token budgets — necessarily, the volume of AI processing energy a developer is allowed to eat — have turn out to be a badge of honor amongst Silicon Valley builders, however that’s an excessively bizarre option to take into accounts productiveness. Measuring an enter to the method makes little sense whilst you possibly care extra in regards to the output. It could make sense if you happen to’re looking to inspire extra AI adoption (or promoting tokens), however now not if you happen to’re looking to turn out to be extra environment friendly.

Imagine the proof from a brand new magnificence of businesses working within the “developer productiveness perception” area. They’re discovering that builders the use of gear like Claude Code, Cursor, and Codex generate much more approved code than they did earlier than. However in addition they in finding that engineers have to go back to revise that approved code way more incessantly than earlier than, undercutting claims of larger productiveness.

Alex Circei, the CEO and founding father of Waydev, is development an intelligence layer to trace those dynamics; his company works with 50 other consumers that make use of greater than 10,000 instrument engineers. (Circei has contributed to TechCrunch previously, however this reporter had by no means met him earlier than.)

He says that engineering managers are seeing code acceptance charges of 80% to 90% — that means the percentage of AI-generated code that builders approve and stay — however they’re lacking the churn that occurs when engineers need to revise that code within the following weeks, which drives the real-world acceptance fee down between 10% and 30% of generated code.

The upward push of AI coding gear led Waydev, based in 2017 to supply developer analytics, to completely transform its platform within the ultimate six months to deal with the proliferation of speedy coding gear. Now, the corporate is freeing new gear that observe the metadata generated through AI brokers, providing analytics at the high quality and value in their code to supply engineering managers with extra perception into each AI adoption and efficacy.

Whilst analytics corporations have an incentive to focus on the issues they in finding, the proof is mounting that enormous organizations are nonetheless working out the way to use AI gear successfully. Primary corporations are noticing — Atlassian got DX, every other engineering intelligence startup, for $1 billion ultimate yr, to lend a hand its consumers perceive the go back on funding on coding brokers.

The information from around the trade tells a constant tale: Extra code is being written, however a disproportionate quantity of it isn’t sticking.

GitClear, every other corporate on this area, revealed a file in January that discovered AI gear larger productiveness, but additionally that its information confirmed “common AI customers averaged 9.4x upper code churn than their non-AI opposite numbers” — greater than double the productiveness features the gear equipped.

Faros AI, an engineering analytics platform, drew on two years of purchaser information for its March 2026 file. The discovering: code churn — strains of code deleted as opposed to strains added — had larger 861% beneath top AI adoption.

Jellyfish, which expenses itself as an intelligence platform for AI-integrated engineering, gathered information on 7,548 engineers within the first quarter of 2026. The company discovered that the engineers with the biggest token budgets produced probably the most pull requests (proposed adjustments to a shared codebase), however the productiveness growth didn’t scale. They completed two occasions the throughput at 10 occasions the price of tokens. In different phrases, the gear are producing quantity, now not worth.

A lot of these statistics ring true whilst you communicate to builders, who’re discovering that code assessment and technical debt are stacking up, at the same time as they revel within the freedom of the brand new gear. One commonplace discovering is the variation between senior and junior engineers, with the latter accepting way more AI-generated code, and coping with a bigger quantity of rewriting as a result.

Nonetheless, at the same time as builders paintings to know precisely what their brokers are as much as, they don’t watch for turning again anytime quickly.

“It is a new technology of instrument building, and you have got to evolve, and you’re pressured to evolve as an organization,” Circei informed TechCrunch. “It’s now not like it’ll be a cycle that can move.”