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

There’s an previous noticed in administration: What you measure issues. And, sometimes, you get extra of no matter you’re measuring.

Software program 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 should be measuring is much less clear.

Huge token budgets — primarily, the quantity of AI processing energy a developer is permitted to devour — have grow to be a badge of honor amongst Silicon Valley builders, however that’s a really bizarre means to consider productiveness. Measuring an enter to the method makes little sense while you presumably care extra in regards to the output. It would make sense when you’re making an attempt to encourage extra AI adoption (or promoting tokens), however not when you’re making an attempt to grow to be extra environment friendly.

Think about the proof from a brand new class of corporations working within the “developer productiveness perception” house. They’re discovering that builders utilizing instruments like Claude Code, Cursor, and Codex generate much more accepted code than they did earlier than. However in addition they discover that engineers should return to revise that accepted code way more typically than earlier than, undercutting claims of elevated productiveness.

Alex Circei, the CEO and founding father of Waydev, is constructing an intelligence layer to trace these dynamics; his agency works with 50 completely different clients that make use of greater than 10,000 software program engineers. (Circei has contributed to TechCrunch prior to now, 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% — which means the share of AI-generated code that builders approve and preserve — however they’re lacking the churn that occurs when engineers should revise that code within the following weeks, which drives the real-world acceptance charge down between 10% and 30% of generated code.

The rise of AI coding instruments led Waydev, based in 2017 to offer developer analytics, to completely rework its platform within the final six months to deal with the proliferation of fast coding instruments. Now, the corporate is releasing new instruments that monitor the metadata generated by AI brokers, providing analytics on the standard and value of their code to offer engineering managers with extra perception into each AI adoption and efficacy.

Whereas analytics corporations have an incentive to focus on the issues they discover, the proof is mounting that enormous organizations are nonetheless determining find out how to use AI instruments effectively. Main corporations are noticing — Atlassian acquired DX, one other engineering intelligence startup, for $1 billion final yr, to assist its clients perceive the return on funding on coding brokers.

The info from throughout the business tells a constant story: Extra code is being written, however a disproportionate quantity of it isn’t sticking.

GitClear, one other firm on this house, published a report in January that discovered AI instruments elevated productiveness, but additionally that its information confirmed “common AI customers averaged 9.4x larger code churn than their non-AI counterparts” — greater than double the productiveness positive aspects the instruments supplied.

Faros AI, an engineering analytics platform, drew on two years of buyer information for its March 2026 report. The discovering: code churn — strains of code deleted versus strains added — had elevated 861% underneath excessive AI adoption.

Jellyfish, which payments itself as an intelligence platform for AI-integrated engineering, collected data on 7,548 engineers within the first quarter of 2026. The agency discovered that the engineers with the biggest token budgets produced probably the most pull requests (proposed modifications to a shared codebase), however the productiveness enchancment didn’t scale. They achieved two instances the throughput at 10 instances the price of tokens. In different phrases, the instruments are producing quantity, not worth.

These sorts of statistics ring true while you speak to builders, who’re discovering that code overview and technical debt are stacking up, whilst they revel within the freedom of the brand new instruments. One widespread discovering is the distinction between senior and junior engineers, with the latter accepting way more AI-generated code, and coping with a bigger quantity of rewriting as a consequence.

Nonetheless, whilst builders work to know precisely what their brokers are as much as, they don’t anticipate turning again anytime quickly.

“This can be a new period of software program growth, and it’s a must to adapt, and you’re compelled to adapt as an organization,” Circei instructed TechCrunch. “It’s not like will probably be a cycle that can move.”