I was watching a YouTube ad for an app I still haven’t bothered to open. It shows someone at a bookshop, or a library, book in hand, actually reading it, when someone cuts in to tell them, pretty assertively, that they’re doing it wrong. Just take out your phone, scan the book, and let the app read it to you instead.
AI-powered, probably, though I couldn’t tell you the details because the pitch itself was enough to turn my stomach.
Reading is one of the most basic things school ever asks you to master. Telling people they can graduate without needing to do it themselves isn’t efficiency. It’s handing off a cognitive skill that took years to build for a task that takes minutes to do yourself.
That ad is what got me thinking about this whole conversation differently.
You’ve probably scrolled past a version of it yourself—TikTok explainers, Instagram reels, news segments about protests—all circling the same idea that AI is a wrecking ball coming for every job.
And I get why that narrative sticks. It’s simple, it’s scary, and simple-and-scary travels fast.
But it’s not actually what’s happening. Not entirely.
The real story is messier, and honestly more interesting, because it’s not just about AI replacing people but about people quietly losing the ability to solve problems without it.
At the exact same time, the economics of running that AI are getting so expensive that in a growing number of cases, paying a human is now the cheaper option.
Two forces, pointing in directions nobody predicted a few years ago.
That’s the Token Trap. Let’s get into it 👇
AI Dependency vs. AI Replacement: The Shift Nobody’s Talking About
AI dependency is what happens when “let AI handle it” becomes your only mode, and the underlying skill quietly atrophies because you never have to use it.
The thing about the “AI is taking our jobs” narrative is that it treats views things with that black-and-white lens hardly anyone’s fond of 😑
Either AI does the task, or a human does. But real life is mostly the grey area in between, and that’s exactly where the actual damage is happening.
The deeper risk isn’t a layoff notice. It’s what happens when “AI-native,” meaning you reach for a model before you reach for your own brain, becomes the only way someone knows how to work.
Take search as an example. A generation is growing up that considers “let me look that up” to mean asking one model and accepting whatever comes back. Forget the concept of opening ten tabs and cross-referencing sources.
Reading a book gets replaced by having it read to you.
Debugging gets replaced by pasting an error into a chat window and waiting.
Sound familiar? 😬
Don’t get me wrong, none of these are “wrong” on their own. I use AI for research, for scaffolding code, for getting unstuck fast.
However, there’s a real difference between using a tool to move faster and losing the underlying skill entirely because you never have to exercise it. One is augmentation. The other is dependency.
Dependency is a lot harder to notice from the inside. You don’t feel the skill leaving; you just gradually stop reaching for it.
Note: This isn’t an argument against using AI. It’s an argument for knowing why you’re using it and what you’d do if it wasn’t there.
And now it’s becoming clear that the same dependency is running head-first into a money problem. The tools we’re leaning on the hardest are turning into some of the priciest bills companies have ever had to pay.
Welcome to the Token Trap, where we have dependency and runaway costs, hitting at the exact same time, with neither one small enough to ignore.
The “Smart Middle”: Why AI Token Maxing at Work Is a Red Flag
There’s a specific type of engineer I’ve learned to watch out for, and it’s not the one who avoids AI. It’s the one who burns through their entire monthly AI quota in two days and treats it like a badge of honor.
That’s not a flex. That’s someone throwing garbage at a model and hoping the output holds together well enough that nobody has to fix it later.
Except someone always does, and it’s usually not them 😒
On the other end, there’s the engineer who refuses to touch AI tools at all, on principle. I get the instinct, but at this point that’s its own kind of falling behind. These tools do speed up real work when you use them with intent.
What I actually want to work with, what I think most teams need, is the smart middle: people who use AI to make an existing workflow faster, or to ideate something new, rather than reaching for it because reaching for it is the culture.
Not maximizing usage. Maximizing outcome per dollar spent.
Why Nvidia’s CEO Wants Engineers Spending $250K on AI Tokens
Part of why “more tokens” got treated as a virtue in the first place traces back to a specific, very public data point. Nvidia CEO Jensen Huang said on the All-In Podcast that if one of his $500,000-a-year engineers only spent $5,000 on AI tokens annually, he’d be, in his words, “deeply alarmed.”
His stated expectation was that a $500,000 engineer should be consuming at least $250,000 worth of tokens a year. That’s roughly half their salary, spent on inference.
Then he compared skipping AI tools to a chip designer insisting on paper and pencil instead of CAD software.
I understand the argument he’s making (even though the numbers alone made my jaw slightly droop), but Huang runs a company that sells the chips those tokens run on. Nvidia itself has reportedly been trying to spend $1–2 billion a year on tokens for its own engineering team. That alone means his incentives, and yours, are not the same incentives.
What Happened When Uber and Microsoft Actually Tested That Advice
Unsurprisingly, the companies actually paying these bills are starting to push back hard on that framing (after, you know, they fell prey to it).
Uber burned through its entire 2026 AI coding tools budget in just four months. Mind you, a budget the company had built after actively incentivizing employees to use more AI through an internal leaderboard ranking teams by usage.
Think about that for a second. They built a leaderboard to encourage more AI usage, then the bill showed up.
Adoption at Uber climbed from 32% of engineers in February to 84% by March, and monthly cost per engineer ranged from $150 to $250 on average, with power users running between $500 and $2,000. Uber’s own CTO reported spending $1,200 in a single two-hour session during a personal demo 🙈
Microsoft went a different direction entirely by reportedly canceling most of its direct Claude Code licenses, moving engineers toward GitHub Copilot CLI instead.
Note 👀
Neither company failed to adopt AI. These are companies that adopted it aggressively, watched the bill arrive, and are now recalibrating toward outcomes instead of raw usage volume.
And to close the loop, consider that Uber capped employee AI spending at $1,500 per tool per month after the budget blew up, and an engineer running two tools at that cap would cost roughly $36,000 a year in tokens alone which is about 11% of a typical Uber software engineer’s $330,000 total pay. That’s not AI quietly saving money in the background, but AI showing up as a second line item on top of the human salary it was supposed to be replacing.
At high enough usage, on high-stakes or complex tasks, the “replacement” doesn’t replace the cost of the person—it stacks on top of it.
Explore: You Need To Work Smarter, Not Harder, With AI
Thinking Tokens: The Hidden AI Cost Behind Overthinking
Thinking tokens are the internal reasoning steps a model generates before it writes the answer you actually see. Most providers bill for them at the same rate as visible output, even though you never read a word of it.
People using these tools have never even seen a line item for this. Here’s where it actually costs you.
What Thinking Tokens Actually Are and Why They’re So Expensive
When you ask a reasoning model a question, it doesn’t jump straight to an answer. It first generates an internal chain of reasoning, weighing approaches and checking its own work, before it writes what you actually see.
On complex, thinking-heavy tasks, that hidden layer can end up as the majority of a request’s total spend, with some workloads landing in the 50–70% range once you actually break down the bill line by line.
Independent industry estimates put the multiplier at 10 to 30 times the visible output cost for complex tasks, and thinking tokens are billed at the same output rate as the response you actually read, whether or not the provider ever shows them to you.
What the maths mean:
- On GPT-5.4, a request with zero reasoning effort might cost $0.0166
- The same prompt, with high reasoning effort (430 thinking tokens instead of 0), hits $0.0331
Same prompt, same model, and the cost doubled purely because the model spent more tokens “thinking” before it answered.
The Overthinking Problem: Why More Reasoning Doesn’t Mean Better Answers
What should actually bother you is that more thinking doesn’t reliably mean a better answer.
A 2025 empirical study on reasoning length and correctness found that accuracy in LLMs often plateaus, and can even decline, once reasoning length crosses a certain threshold.
Note 👇
The same research found that models tend to overthink simple problems, generating unnecessarily long outputs, while underthinking genuinely hard ones. This means the model is misjudging problem difficulty and failing to calibrate how much reasoning a task actually needs.
So you’re not just paying more for deeper thinking.
On easy problems, you’re frequently paying a premium for the model to talk itself into circles, introducing noise that a person, or a simple deterministic script, wouldn’t have introduced in the first place.
The Cheap Model Fallacy: Claude Sonnet 5 vs. Opus 4.8
What if I told you that “cheaper” models don’t automatically mean cheaper bills in agentic workflows?
Note: An agentic workflow is when you chain a model through multiple steps, tool calls, and decisions in a sequence, rather than a single prompt-and-response. Check out my updated portfolio’s AI assistant Eve that is now agentic by introducing tools to parse and infer role fit.
Claude Sonnet 5 lists at a fraction of Opus 4.8’s price on paper. But Sonnet 5 generates roughly 30% more tokens than Opus for comparable tasks, and in agentic workflows with multiple steps and tool calls, that verbosity compounds.
Add adaptive thinking running by default. Add a newer tokenizer that counts more tokens for the same text. Now an output-heavy agentic workload can end up costing the same as Opus 4.8, or more.
The sticker price says Sonnet is the budget option, but the actual bill, after enough steps and tool calls, can say otherwise.
At the end, optimizing for the price on the model card instead of total tokens burned per completed task is the trap hiding inside the trap.
Tip 👀
If you’re building anything on top of an LLM API, check whether your provider exposes a reasoning effort or thinking budget parameter (Anthropic’sthinkingconfig, OpenAI’sreasoning.effort). Defaulting to medium or adaptive effort on every single call, including trivial ones, is one of the fastest ways to quietly inflate a bill without noticing.
AI Workflow Files: How to Cut Inference Costs With Deterministic Order
None of this means AI is a bad investment. It means the tools need structure, or the costs run away from you.
Context Reuse: How Workflow Files Cut a Bill From $11.55 to $1.06
Workflow files solve the core problem of context accumulation. That’s what happens when every new session starts cold, and the model has to “re-learn” your entire project from scratch.
Your conventions, file structure, preferences, naming patterns. You’re paying input-token cost for that re-education every single time, even though nothing has actually changed.
Think of it like hiring a contractor who forgets everything about your house each morning and needs a full briefing before they pick up a tool.
Storing your rules in something like .cursorrules or a SKILL.md, AGENTS.md, or CLAUDE.mdfile fixes this (different names, same idea). Instead of re-explaining your project in every prompt, the model reads a compact reference file and picks up where you left off.
I’ve seen a single session’s cost drop from $11.55 to $1.06 just by doing this with the same task and output but a fraction of the tokens because the model wasn’t re-deriving context it already had in front of it.
Related: All You Need To Know About AI Workflow Files And How To Use Them
The other half of it is knowing what shouldn’t go to a model at all.
Not Every Task Needs a Model: When to Use Deterministic Code Instead
Not every task needs a “brain” behind it. Copying text from a screenshot, reformatting JSON, renaming a batch of files by pattern, or parsing PDFs are mechanical tasks with one correct answer, every time.
It’s what developers call deterministic since you have the same input always producing the same output, no guessing involved. For example, if you can describe the task as “convert X to Y” or “do Z to every file in this folder,” it probably belongs in a script instead of a chat window.
Push those tasks to the actual app layer: code, scripts, regex, whatever handles them reliably. That layer is free to run, doesn’t hallucinate, and never has a bad day.
Tip 🔥
You can couple these deterministic tools in the agentic chain before involving AI. This way you provide structured, consistent data that AI models can better infer from.
Reasoning models are for judgment calls. Deterministic code is for everything else 💁♀️
AI Slop vs. Human Judgment: The Social Contract at Risk
There’s a version of this conversation that doesn’t focus on money but on how people’s working relationships change with the tools they use and with each other.
Reading something clearly AI-generated can be tiring, not because the sentences are technically incorrect (they’re often okay), but because the writing lacks any real purpose. No person actually decided this word mattered more than that one. You can feel the absence of a decision, even when you can’t point to what’s broken.
You know that feeling when you get a customer support email that’s clearly been written by a bot? The one that addresses your exact complaint but somehow doesn’t actually answer anything? That’s output without judgment (aka AI slop). It exists, it’s grammatically complete, and utterly useless.
That same exhaustion appears at work when AI-for-everything gets mandated without asking what problem it’s solving.
What a person brings to any partnership with a tool is judgment by knowing when to reach for it and, just as often, when not to. If you strip that out, but haven’t made the person more productive, you’ve made them a worse version of the tool they’re being told to use.
Which loops back to that reading app from the opening. If we hand off reading, researching, and writing—the actual skills that make a person’s contribution worth something in the first place—instead of becoming partners with these tools, we become bill-payers for a system we don’t quite grasp enough to challenge.
That’s not a hypothetical. That’s the literal direction the “just let AI handle it” framing points.
It’s a Wrap
If there’s one shift worth making after all of this, it’s to stop measuring AI use by tokens spent and start measuring it by outcomes. By dollars per resolved ticket, per accepted PR, per feature that actually shipped to a real user.
Uber’s own leadership said as much. Their COO noted it’s difficult to draw a direct line between rising AI tool usage and new features actually reaching users, which is exactly the gap that a cost-per-outcome lens exposes.
Token volume was never the real scoreboard. It just looked like one because it was the easiest number to track.
AI shouldn’t be your most expensive hire 🙅♀️
Escaping the Token Trap means moving toward setting up conditions so the model doesn’t waste tokens figuring out what you already know. This can be:
- Workflow files that store your project context
- Deterministic scripts that handle the mechanical stuff
- Reasoning effort tuned to task difficulty, not left on “high” because nobody checked the default
Use AI like a forklift for the mind by letting it handle the heavy, repetitive lifting so you can focus on the work that actually needs you. But a forklift doesn’t replace the person operating it, and it definitely doesn’t replace the judgment about what needs moving in the first place.
Skip that part, and you’re not augmenting anything. You’re, in the words of something I recently heard that captures it well, paying a Ferrari’s lease to do a bicycle’s job.
Thanks for reading this one; it’s been on my mind for a while.
I’ll catch ya next time.
Keep thinking!
