This is the kind of idea that sounds responsible and ends up making people act weird.
Cognition is rolling out a system that tries to stop “tokenmaxxing” by grading engineers on “AI efficiency.” Use the AI well, and your spending limits go up. Use it badly, and your limits get tightened. They’re also tying this grading into Devin Coach, their AI assistant, so the tool can give engineers targeted feedback on how to improve.
On paper, sure. Everybody wants to cut waste. AI usage costs money. If people are spamming prompts, looping the model, or using it like a slot machine until something decent falls out, that bill gets ugly fast. And in a lot of companies, budgets are blunt instruments. You get a fixed allowance, you burn it down, and somehow the only lesson learned is “use it all before they cut you next quarter.”
So I understand the itch Cognition is scratching. They’re trying to swap “everyone gets the same bucket” for “prove you can drive, and we’ll give you a faster car.”
But grading engineers on AI efficiency is also a great way to accidentally reward the wrong behavior.
Because the thing you’re measuring isn’t “good engineering.” It’s “how cheaply you can get the AI to produce something that looks like progress.” Those are not the same. Sometimes the difference is the whole job.
Imagine you’re working on a tricky bug that only happens in a rare edge case. The fastest way might be to throw a bunch of careful context at the AI, explore a few angles, and iterate until you find the pattern. That might be “token-inefficient.” It might also be the difference between a fix and a patch that breaks again next week.
Now imagine you know your spending limit can shrink if you look inefficient. What do you do? You stop exploring. You stop asking “dumb” questions. You avoid using the AI on the hard stuff and save it for easy wins where you can look clean and efficient. Or you squeeze prompts until they’re too short, too vague, and the AI fills in the blanks with confident nonsense. Your token bill improves. Your software gets worse.
That’s the tension: companies want to pay for outcomes, but they’ll end up paying for whatever their measurement system calls an “outcome.”
There’s also a culture problem hiding in here. “Tokenmaxxing” is a funny word, but the behavior it points to is real: people will game any system that controls their resources. If the system says “be efficient,” people will learn how to look efficient. Not how to be effective.
Say you’re a junior engineer. You’re still learning how to ask good questions. You’ll naturally burn more tokens. Under a grading system, you’re now expensive by definition. Even if you’re smart. Even if you’re improving fast. If your limit gets tightened, you get less practice, less feedback, and more pressure. That’s how you turn a learning curve into a dead end.
Meanwhile the senior engineer who already knows the codebase can be “efficient” because they don’t need to ask as much. Great for them. But the system might quietly widen the gap between the people who have context and the people who need to build it.
And then there’s Devin Coach being used to “help.” I’m not against coaching tools. But I don’t love the vibe of “the same system that grades you also teaches you.” That can be supportive, or it can feel like a manager sitting behind your eyes. If the coaching is aligned with the grading metric, it will nudge people toward what scores well, not what ships safely.
Picture a real-world moment: a team is two days from a deadline. An engineer has to decide whether to use the AI to generate a test suite they can review carefully, or to skip the tests and just patch the bug. If “efficiency” gets rewarded in a simplistic way, the system might push them toward the option that uses fewer tokens, not the option that reduces risk. And when that blows up in production, the tokens won’t be the expensive part.
To be fair, there’s a version of this that could be genuinely healthy. If the grading is thoughtful, if it accounts for task difficulty, if it rewards solid output instead of cheap output, and if it’s used as guidance rather than punishment. If it’s more like “your prompts are unclear, here’s how to fix that” and less like “your budget is cut, good luck.” Some people do waste tokens out of habit, not need. A gentle constraint can force better habits.
But once you connect a score to a limit, it stops being gentle. It becomes power. And power changes behavior fast.
The bigger picture is that enterprises are shifting away from fixed token budgets to performance-linked adjustments. I get why. Finance teams hate open-ended usage. Leaders want levers they can pull. But the moment you turn AI usage into a personal performance metric, you’ve turned a tool into a test. And tests reshape work in ways nobody admits until it’s too late.
Maybe Cognition nails this. Maybe it actually reduces waste without turning engineering into a weird game of prompt minimalism. I just don’t think you get that for free. You have to choose what you’re optimizing for, and you have to accept the trade-offs.
So the real question is: should companies treat AI usage like an expense to control per person, or like shared infrastructure where the only thing that matters is whether the work got better?