10x the Tokens, 2x the Value: When Using AI Becomes the Goal

A few months ago I started seeing the same odd story repeat across some of the biggest tech companies in the world.

It always starts the same way. Leadership gets excited and tells engineers to use AI everywhere. Write code with AI. Write docs with AI. Do your research with AI. Some teams take it a step further and build internal leaderboards that rank who uses AI the most.

For a while, it looks great. The productivity charts go up. Work ships faster. Everyone is happy.

Then a few months pass, and someone finally looks at the bill.

And then the bill shows up

This is not a made-up example. The receipts are public now.

At Meta, an employee built an internal leaderboard nicknamed "Claudeonomics." It ranked staff by how many tokens their AI agents used. People earned badges and titles like "Model Connoisseur" and "Cache Wizard." You can guess what happened next. Engineers started competing. They wrote longer prompts and ran several agents at once just to climb the board. The top user reportedly averaged around 281 billion tokens. The leaderboard was quietly taken down about two days after it went live. (Inc., Tom's Hardware)

At Amazon, internal targets pushed for more than 80% of developers to use AI tools every week, and that usage showed up on leaderboards too. Some employees admitted they ran agents on pointless tasks just to pump up their numbers. One worker described "so much pressure to use these tools." Another called it "perverse incentives." Amazon later backed away from raw token counts and switched to a metric it calls "normalized deployments." (Tom's Hardware, AI Magazine)

And at Uber, the numbers got loud. CTO Praveen Neppalli Naga told The Information that the company had burned through its entire 2026 AI coding budget in just four months. In his words: "I'm back to the drawing board because the budget I thought I would need is blown away already." Around 5,000 engineers were using the tools, and some monthly bills ran from $500 to $2,000 per person. Neppalli Naga reportedly spent $1,200 in tokens during a single two-hour demo. Uber ended up capping spending at $1,500 per person per month. Its COO, Andrew Macdonald, was honest about the payoff. The link between rising Claude Code usage and real customer value, he said, "is not there yet." (Fortune, TechCrunch)

Higher cost is fine. The ratio is the problem.

Spending more is not a problem on its own. It is only a problem if the value does not go up at the same rate.

So here is the real question. If you spend 10x the tokens, do you get 10x the value?

The data says no.

The engineering analytics firm Jellyfish found that the engineers who used the most tokens were about twice as productive as lighter AI users. But they spent roughly 10x the tokens to get there. Two times the output at ten times the cost. That is volume, not value. (TechCrunch)

A bigger study from Faros AI looked at more than 10,000 developers across 1,255 teams over as long as two years, and it found the same split from another angle. The output clearly went up. Developers completed 21% more tasks and merged 98% more pull requests. But the quality moved the wrong way. There were 9% more bugs per developer, pull requests grew 154% larger, and review times got 91% longer. The most important finding is also the quietest one: "No significant correlation between AI adoption and improvements at the company level." The work got faster, but the business results did not follow. (Faros AI)

There is one more study worth sitting with. METR asked 16 experienced open-source developers to finish real tasks, some with AI and some without. They expected AI to make them about 24% faster. It actually made them 19% slower. And even after they were done, they still believed it had sped them up by 20%. The gap between feeling productive and actually being productive turned out to be huge. (METR)

This is an old bug, not a new one

If this pattern feels familiar, it should. We have seen it in software for a long time. It even has a name: Goodhart's Law, which says that when a measure becomes a target, it stops being a good measure.

The moment "token usage" became the thing leadership watched, "token usage" became the thing people chased. Slowly the belief took hold that more tokens meant more work. Before long, teams were improving their AI usage instead of their business results. They were tuning the number on the dashboard, not the outcome it was supposed to stand for. And as it turned out, the heaviest token users were not reliably the most valuable engineers.

That is the whole trap. AI is not the villain here. The real problem is blindly leaning on it.

AI is a tool. Treat it like one.

A tool exists to make a person more capable. A tool is not there to do your thinking for you.

So the people who do best in the AI era will not be the ones who use the most AI. They will be the ones who know where to use it and where to trust their own judgment instead. Knowing when to reach for the agent, and when to close the laptop and think, is becoming the real skill.

That is why I no longer believe the most important skill of this era is prompt engineering. I think it is three quieter ones:

• Cost awareness. Every token is a real dollar, and "more" is never free.
• Critical thinking. Check whether the output is actually correct and actually useful, not just convincing.
• Good judgment. Know which problems deserve AI, and which ones deserve you.

Because in the end, tokens were never the point.

Value is.

Sources

• Inc. What Is 'Tokenmaxxing'? (Meta "Claudeonomics" leaderboard)
• Tom's Hardware Big Tech has a 'tokenmaxxing' habit
• AI Magazine Why Amazon dropped its internal AI usage leaderboard
• Fortune Uber burned through its entire 2026 AI budget in four months
• TechCrunch The token bill comes due (Jellyfish 2x for 10x finding)
• Faros AI The AI Productivity Paradox
• METR Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity

---

Md. Tausif Hossain leads engineering at DevTechGuru and runs TechnicalBind, an independent software studio. He writes about AI-assisted engineering, distributed teams, and the craft of shipping. Reach him at tausif.bd or @tausif1337.