Tokenmaxxing is dead, modelmaxxing is next. Uber learned that in 4 months, I learned it in a week.
My AI Quota Died in 20 Minutes. A 19th-Century Coal Paradox Explains Why.
Last Wednesday, I open my quota dashboard and 75% of my monthly allowance is gone. 1 task. Cleaning up a checkout flow with Ultracode, Claude Fable 5's coding mode. I've written about that specific disaster before, in I Found Claude Fable 5's Best Use Case. It Cost 75% of My Monthly Quota to Get There., so I won't replay the autopsy here.
What I haven't told you yet is what happened next.
Same week, next day, I switch to Sonnet. And I do the exact opposite of what a sane person would do after watching 75% of a budget disappear in 20 minutes. I max it out for everything, a variable rename, a typo fix in a README, tasks that would embarrass a Haiku call, I'm sending to the model with the biggest context window and the highest price tag, because that's the default and defaults are what you reach for when you're annoyed. 2 days later, the week's quota is dead too, and the dashboard doesn't flash YOU DIED in giant red letters, but it might as well.
Same reflex, 2 days apart, dressed differently each time. I'm not explaining it yet. I'm just putting the fact on the table: I pay for a subscription specifically so I stop counting tokens, and I still hit 0 before the week is half over.
Tokenmaxxing is dead. Modelmaxxing is next, and there's a 160-year-old reason why.
The Coal Question
There's a 160-year-old paper that explains exactly what I did to myself, and it has nothing to do with AI.
In 1865, an economist named William Stanley Jevons published The Coal Question. His worry was that England was going to run out of coal, and that James Watt's improved steam engine, the one that made coal-burning dramatically more efficient, would actually make things worse. Turns out he had it backwards in the most instructive way possible. The efficient engine didn't reduce coal consumption. It exploded it.
Coal got cheaper to use per unit of output, so it got used everywhere, in more machines, for more purposes, until total consumption dwarfed anything the inefficient engines had ever burned. Efficiency didn't save the resource. It made the resource feel free, right up until it wasn't.
This isn't a dusty economics factoid I'm dragging in for flavor. Satya Nadella cited Jevons paradox by name on social media in early 2025, right after DeepSeek's release, when AI circles were debating whether cheaper inference would shrink AI spending. His read, paraphrased: don't count on it, cheaper compute means more compute gets burned, not less. He was right, and I am living proof, 1 checkout page and 1 README typo at a time.
A flat-rate AI subscription runs the exact same trick coal did in 1865. It doesn't feel like a resource with a bottom. It feels infinite, right up until the dashboard tells you otherwise, usually mid-task, with zero warning.
The Company-Sized Version of My Mistake
I'm not the only one who ran this experiment. I'm just the smallest.
Uber burned through its entire 2026 AI budget in about 4 months. Not the quarter's budget, the year's. The company's response wasn't a strategy memo, it was a hard cap: $1,500 per employee per month, per tool, no exceptions. Meta had its own version of the same story, except theirs had a leaderboard. Internally nicknamed "Claudeonomics," it let 85,000 employees compete on raw token consumption, gamified, visible, encouraged. 1 employee hit 281 billion tokens in a single 30-day stretch, a number I've reread 3 times and it still doesn't parse as a human workload. The leaderboard got pulled quietly once finance actually looked at the invoice.
Random tangent, not going anywhere useful: my own quota dashboard rounds to the nearest 5%, so I found out about the 75% burn from an automated email, not the UI itself. Half these platforms still treat "your bill is on fire" as a batch job instead of a push notification. Anyway.
Different company, same arc as my week: cheap access, no friction, maximum consumption, until someone with visibility into the actual cost pulls the plug. I felt an echo of that same instinct reading about the metric that makes Anthropic's own team sound the alarm, which tells you the discipline problem doesn't stop at the vendor's front door either.
Uber has a CFO, and so does Meta, someone in that org chart eventually looks at a spreadsheet and says stop. I don't have that person. Nobody's watching my burn rate except me, and I clearly wasn't watching closely enough, or I'd have caught the Sonnet spiral before it caught me.
The Turn Toward Modelmaxxing
The fix that's emerging at company scale isn't "use AI less." It's "use the expensive model less."
Coinbase is the cleanest example I've found. CEO Brian Armstrong has kept the company's AI costs roughly flat even as internal usage climbed exponentially, and the mechanism isn't a hiring freeze on prompts, it's routing. Cheap model when the task doesn't need more, expensive model when it genuinely does. That's the whole idea behind what people are now calling modelmaxxing: not maximizing token spend, maximizing the match between task complexity and model tier.
The sharpest single case I found is Lindy, a startup that moved 100% of certain traffic categories from Claude to DeepSeek. Inference costs on those routes dropped by roughly 90%, and on several core use cases, performance actually got better, not worse. That part should sting a little if you've been defaulting to the priciest model out of habit rather than need. Routing isn't automatically a downgrade. Sometimes the expensive model was never the right tool, it was just the reflexive one.
2 more data points, worth a sentence each and no more. Salesforce's projected 2026 Anthropic bill sits around $300 million, and CEO Marc Benioff has publicly said he wants a smart router that decides which queries actually need a frontier model. At Bold Metrics, CTO Morgan Linton now dictates model choice directly to his engineering team rather than leaving it to individual habit. Different scales, same conclusion: somebody decided the default couldn't stay the default anymore.
I Already Had Half the Answer
Here's the part that stings in a different way. I'd already solved a piece of this, months before it had a name.
Back in May, I published 94% of My Claude Code Tokens Went to the Wrong Model. So I Stopped Paying Opus to Do Haiku's Job., a breakdown showing that 94% of my Claude Code tokens were going to the wrong model, Opus doing work that Haiku could have handled without breaking a sweat. At the time it felt like a personal audit, an optimization I figured nobody else would care about. It wasn't a movement. It was just me, annoyed at a bill, digging through logs.
That same instinct is now an industry practice with a name, case studies, and CEOs quoting it in earnings calls. The solo dev bricolaging his own routing in a spreadsheet was right before it was fashionable. I just never gave it a name, and I stopped applying it consistently, which is exactly how I ended up back here with a dead Sonnet quota 2 days after a dead Fable quota.
Routing Without a FinOps Team
You don't need Coinbase's engineering org or Salesforce's budget to do this. You need 1 habit, repeated until it's boring.
Before launching an agent, scope the task first and ask, honestly, what tier of model it actually deserves. Not what you're used to reaching for. What the task, stripped of your reflexes, requires. Most tasks don't need the boss-fight model, they need the one that handles a fetch quest without complaint. This is the same discipline that sits underneath the Blueprint method in Vibe Coding, For Real: a task that's badly scoped costs you money no matter which model you throw at it, expensive or cheap. Scoping first is what turns "which model" from a guess into a decision, and it's part of why I still lean on a CLI-first setup instead of stacking MCP servers, less overhead per call means the scoping decision actually shows up in the bill.
This is the manual version of what Jevons paradox is warning you about, and it's the one lever a solo dev actually has. Coal never ran out because efficiency made people careless with it. Tokens work the same way under a flat subscription, it hides the scarcity until the scarcity reasserts itself all at once, at the worst time, with no warning shot. Scoping before you launch is how you see the rarity coming instead of discovering it at 0.
Maybe I'm wrong that this scales past a solo dev or a small team without real tooling behind it eventually. But as a discipline you can start today with zero infrastructure, it works, mostly.
I say mostly because I got burned twice in the same week before I actually corrected course. The habit doesn't eliminate the mistake. It just makes the mistake smaller, and rarer, and a lot less likely to cost you a week's quota over a README typo.
I burned a week's worth of tokens on a checkout page and a typo. Not the AI apocalypse, just me being cheap with attention and expensive with everything else.
Scope the task. The model can wait.
You pay for unlimited tokens so you stop counting them, then you hit zero anyway. The demo vs product checklist in the welcome kit shows you the rate-limiting rule that keeps this from killing your budget in production.