For about six months, the smartest career move inside some of the biggest tech companies was tokenmaxxing, essentially wasting money on purpose.
If you burned the most tokens, you topped the leaderboard. If you topped the leaderboard, you looked productive. Nobody stopped to ask whether any of those tokens shipped anything. That was tokenmaxxing, and it’s already collapsing under its own invoice.
The most common perception is that AI spending is collapsing. But honestly, it isn’t. What died is the belief that more tokens means more value. So I want you to walk through what actually changed, why the metric was doomed from the start, and what the smarter teams are doing instead.
Key Takeaways
- Tokenmaxxing meant ranking people and AI agents by token consumption. Volume was treated as proof of productivity.
- The metric broke because of Goodhart’s Law: once token counts became the target, people gamed them.
- The bills got real. Uber blew its 2026 AI budget in just four months. Salesforce is looking at a roughly $300 million annual Anthropic bill.
- AI RoI and AI costs stopped lining up, because spend scaled faster than any measurable output.
- Cutting token spend isn’t always right. For AI-native teams, rationing the best model can kill the edge that pays for it.
What Was Tokenmaxxing? Inside the Hype of Enterprise AI Spending
Tokenmaxxing is the practice of treating token usage as a productivity score. A token is a unit of AI text, roughly three-quarters of a word, and the logic was simple: the more your team burned, the more ‘innovative” it looked.
This wasn’t a fringe idea. It came from the top. Nvidia’s Jensen Huang said in March that he would be deeply alarmed if an engineer earning $500,000 wasn’t spending $250,000 a year on tokens. Meta, Amazon, and OpenAI built internal leaderboards so staff could compete on consumption. Meta’s ran under the nickname Claudeonomics and logged something like 60 trillion tokens in a single month.
If you’ve ever measured developers by lines of code, you already know how this ends. We threw that metric out decades ago because it rewarded bloat. Tokenmaxxing quietly brought it back.
The appeal was that it was easy. Token counts are trivial to track. Actual value is hard to measure. So leadership reached for the number it could see, not the one that mattered.
Goodhart’s Law: Why the Token-Usage Metric Blew Up AI Costs
The metric measured activity, not output. It just shows that you ran a model. But it didn’t tell you whether the result was correct, useful, or shipped.
That isn’t a theoretical gap. Faros AI looked at two years of data across 22,000 developers and found task completion up, but bugs up 54% and code churn up 861% in high-adoption environments. Throughput went up. So did the mess that throughput left behind.
There’s also a quieter problem. Heavy token use often signals a task that was specced badly, looped too many times, or restarted from scratch. A high score can mean someone did great work, or it can mean an agent flailed for an hour. But the number looks the same either way.
The Goodhart’s law problem
Goodhart’s Law says that when a measure becomes a target, it stops being a good measure. Tokenmaxxing is almost a textbook case.
At Amazon, once token counts fed into how people were judged, engineers reportedly spun up agents on pointless tasks just to push their numbers up. The leaderboard stopped tracking productivity and started tracking who was best at feeding the leaderboard. Amazon pulled it down. The behavior was rational. But the metric was the problem.
The Bills That Forced a Rethink at Salesforce, Uber, and Microsoft
The reckoning didn’t arrive as a philosophy debate. It arrived as an invoice.
Per-token prices have been falling for years, but agentic tools flip that math. A single agent can call a model over and over, using up to 1,000 times more tokens than a plain chatbot query. Cheaper tokens, far more of them, and the bill goes up anyway.
This is the Jevons paradox in miniature. When a resource gets cheaper and more efficient, people don’t use less of it. They use much more. Every drop in token price was met with a bigger jump in token appetite, so the per-unit savings never reached the bottom line. The AI economics here are counterintuitive, and plenty of finance teams learned that the hard way.
Here’s where the big names actually landed:
| Company | What happened | Action taken |
| Salesforce | Anthropic bill on track for about $300M this year | Building a smart router, tracking “agentic work units” instead of tokens |
| Uber | Burned its entire 2026 AI budget in four months | Imposed a $1,500 per month cap per coding tool |
| Walmart | Internal AI bills climbing | Set per-employee usage limits |
| Microsoft | Some engineers running up large individual bills | Reined in individual Claude Code spend |
| Meta / Amazon | Leaderboards got gamed | Took the leaderboards down |
The pattern is consistent. Cost moved from a rounding error to a line item finance actually reads. Once AI costs became material, “use as much as possible” stopped being a strategy and started being a liability.
Why AI Costs and ROI Don’t Always Line Up
Spending climbed. But company-wide output didn’t climb with it. That’s the disconnect at the center of the whole story.
To be fair, this thing cuts both ways. A weak link today doesn’t prove the spending is wrong. But it might mean that our measurement tools are immature, and the value is real but invisible to current dashboards. That’s the honest version of the debate.
The productivity J-Curve, explained with electricity.
The best frame I have seen comes from analyst Azeem Azhar, who compares this to factory electrification. When factories first got electric power, they just swapped gas lamps for electric bulbs and changed almost nothing about how they worked.
The real gains came much later, in three rough stages:
- First, replace the old power source (electric lights instead of gas).
- Then, rip out the central steam engine that drove everything through one shaft.
- Finally, wire each machine individually and redesign the whole floor around electricity.
Productivity actually dipped before it jumped, because the early payoff was small and the rewiring was expensive. Most companies using AI today are still at that stage. They bolted a model onto old workflows and are surprised the returns look thin. The curve says returns come after the redesign, not before.
What Replaces Token Maximizing: Routing, Caching, and Workflow Redesign
The replacement has a name now, and is called token-minimizing. The goal flips from burn the most to get the same result for less, and it’s mostly an AI infrastructure problem, not a willpower problem.
The shift is simple to state. Stop sending every request to the most expensive frontier model. Route the easy stuff to cheaper or open-source models, and save the premium model for problems that genuinely need it.
A few moves are doing most of the work:
- Smart routing: Classification, extraction, and summarization go to small models. Hard reasoning goes to the expensive one. Most teams default everything to the flagship model purely because nobody configured tiers.
- Prompt caching: If you resend the same giant system prompt on every call, you’re paying for identical tokens over and over. Caching stable prefixes cuts that waste sharply.
- Trimming context: Agents often carry thousands of tokens of instructions they don’t use for the task in front of them. Load what the task needs, not the whole manual.
- Workflow redesign: This is the J-curve stage most teams skip. Real savings come from rebuilding the process around AI, not pasting AI onto the old one.
None of this is glamorous. It is the AI economics equivalent of fixing the leaks before buying a bigger water tank, and it is where the actual money is.
A concrete version helps. Say a support agent handles 10,000 tickets a day. Routing the tag-and-classify step to a small model and only escalating the actual reply to a frontier model can cut the per-ticket cost by more than half, with no drop a customer would ever notice. Multiply that by a year and the savings stop being a rounding error and start being a headcount.
My Cost-per-Outcome Test Across Model Tiers
I wanted to see the gap myself, so I priced one realistic task in three ways. Summarize and tag a 6,000-word support thread, then draft a reply. The numbers below are an illustrative model using public list prices.
| Approach | Rough cost per task | Quality delta |
| Frontier model for everything | ~$0.18 | Baseline, slightly over-engineered |
| Mid-tier model for everything | ~$0.04 | Near-identical on this task |
| Routed: small model to tag, frontier to draft | ~$0.06 | Best balance, no visible quality loss |
This wasn’t the savings. It was that the cheap setup matched the expensive one on a task this routine. The frontier model wasn’t buying better output here. It was buying a bigger bill.
That’s the whole tokenmaxxing trap in one table. Defaulting to the most powerful model feels safe, but on most enterprise AI work it just inflates cost per outcome for no return.
When Cutting Token Spend Is the Wrong Move
Rationing tokens can backfire. If your edge depends on running the best available model, a cap is just self-sabotage. And some teams know this. At Databricks, an engineering leader said the AI budget for engineers is still effectively unlimited, because they trust their people to use it well.
To be honest, spending caps make the CFO happy and can quietly throttle the experimentation that produced your breakthroughs in the first place. Cutting token spend is good hygiene right up until it becomes the corporate version of skipping lunch to save money and then crashing at 3 p.m.
So the question isn’t, are we spending less? It is: do we know what each token bought? If the answer is yes, heavy spending can be fixed entirely.
But the overcorrection risk is real too. At Accenture, 404 Media reported that leadership started pushing back on staff using AI for trivial chores, like turning PDFs into slides, not long after telling those same employees their promotions depended on adopting AI. Whiplash like that is its own kind of waste.
What This Means for Enterprise AI Budgets in 2026
The core objective is shifting from flat per-person token allowances to outcome-tied spending. Enterprise AI budgets are starting to look less like a buffet and more like any other operating expense, with an owner and a cap.
A few things I would expect to harden over the next year:
- Spend tied to shipped value, with chargebacks to the team and the workflow, not the headcount.
- New success metrics replacing token counts. Salesforce already swapped to agentic work units. Expect more of these.
- A cheap default, premium on request. The expensive model becomes opt-in for hard problems, not the standing default.
The teams that win the agentic era will not be the ones that used the most AI. They will be the ones who knew, line by line, what every token was for.
Final Thoughts
Token volume was never the goal. It was just a proxy that was easy to count and easy to game. But the bills made that obvious.
If you take one thing from this, make it the routing and measurement piece. Pick your success metric before you scale the behavior, keep a cheap model as your default, and reserve the expensive one for work that earns it. Tokenmaxxing is over. Knowing what your AI actually buys is the part that lasts.
FAQ
No. Tokenmaxxing was a metric and an incentive system that treated token consumption as productivity. Total AI spend is a separate budget line, and for most big companies it is still rising. What ended is the belief that more tokens automatically means more value.
It reinforced inappropriate behavior. After counting the number of tokens, people inflated them, the Goodhart’s Law of that kind. Finally, when it became obvious that numbers were being gamed, Meta and Amazon removed their leaderboards.
Cost per outcome. Rather than counting the number of tokens, teams measure the cost to complete a realistic task correctly, the error rate and the business impact per dollar. One early useable version, which Salesforce calls ‘agentic work units,’ is a for-profit idea.
As opposed to getting the same outcome for more tokens, it is getting the same outcome with fewer tokens. In reality, this translates to moving basic tasks to lower cost machines, storing the same context multiple times, and re-engineering processes to eliminate agents from carrying context they don’t use.
Not quite. This is a repricing of expectations, not a retreat from AI. Companies are cutting the spend that does not pay for itself and keeping the spend that does. It does feed the broader nervousness about AI valuations, since flat enterprise growth would dent the providers, but capping wasteful tokens is closer to budget discipline than to collapse.
