The tokenmaxxing trap: why more AI output doesn't mean more productivity

There's a term going around Silicon Valley right now called "tokenmaxxing." It means burning as many AI tokens as possible in a single coding session, treating it like a productivity badge. The more tokens you burn, the harder you're working. The more AI output, the better the outcome.

A TechCrunch investigation published this week is poking some pretty big holes in that idea.

Companies tracking developer productivity across 10,000+ engineers found something uncomfortable: yes, AI tools like Claude Code, Cursor, and Codex generate more code than ever. But engineers are returning to revise that AI-generated code far more often than before. The raw acceptance rate looks fine at 80–90%. But once you track churn — all the rewrites in the weeks that follow — the real rate drops to 10–30%.

More output. More rework. Net gain: unclear.

The input problem nobody's talking about

Here's what I keep coming back to. We've spent years obsessing over AI output: faster models, longer context windows, better reasoning. Almost nobody is asking about the other side of the equation — how you actually talk to the AI in the first place.

Most people are still typing everything. Every prompt, every email, every document, every Slack message. The average person types around 40 words per minute. They speak at 130+. That's a 3x gap hiding in plain sight, before AI even enters the picture.

If you're prompting a coding agent, writing a spec, drafting a client email, or capturing meeting notes — and you're doing all of that by typing — you're leaving a lot on the table.

Voice input changes how you think, not just how fast you work

This is exactly the problem DictaFlow was built to solve. It's an AI-powered dictation app that works system-wide on Mac, Windows, and iOS — any text field, any app. You speak, it transcribes in real time. Your IDE, your browser, your email client, your Slack — all of it.

What surprises most people isn't the speed. It's that they write differently when speaking. More exploratory. Less self-editing mid-thought. Prompts get better because you're not fighting the keyboard while still figuring out what you want to say.

For developers in the tokenmaxxing trap specifically — spending hours crafting elaborate prompts and reviewing output — switching to voice can meaningfully tighten the loop. Describe what you need, hear it back, adjust, iterate. The bottleneck shifts from typing speed to thinking speed.

The actual metric worth tracking

The tools that survive the current AI hype cycle won't be the ones generating the most tokens. They'll be the ones that shrink the gap between a thought in your head and something useful in the world.

Tokenmaxxing optimizes the wrong thing. Volume isn't the point. Iteration speed is.

Voice input is one of the fastest ways to move that needle. If you're already using AI coding tools and still typing every prompt, try dictating for a week. The difference is immediate.