The tech industry is entering a new phase of AI spending discipline, and the latest data from Jellyfish suggests the winners are not the biggest token burners.
Loading audio narration…
In a recent study, engineering intelligence company Jellyfish said the top 10% of Claude Code users consumed about 10 times as many AI tokens as the median developer, but produced only about twice the output.
AI tokens are small chunks of text that AI models break words and inputs into to process them. They are also a way to price AI usage, usually at a cost per million tokens.
Nicholas Arcolano, Jellyfish’s head of AI and research, said that gap is the clearest sign that extreme “tokenmaxxing,” where workers use as many AI tokens as possible, is not a sustainable strategy.
“CFOs are getting on people’s case,” he told me in an interview. “In most companies, you can’t work without showing your receipts. Customers want to move fast, and they’re willing to spend money here, but they can’t do it without showing they’re spending responsibly and having an impact.”
Jellyfish has access to hundreds of companies and the coding behavior of hundreds of thousands of software engineers. It found that consuming a massive number of tokens may not always generate real returns, while increasing costs for companies. This is pushing the tech industry into a new phase in which AI efficiency is becoming more important.
“Even if you rationalize that you’re getting more value from these things than you would get from a human doing the same work, if token costs surge, a CFO will still worry that you broke their spreadsheets,” Arcolano said.
Indeed, the Jellyfish report shows how quickly token use can scale at the high end. Weekly Claude Code consumption for top AI adopters reached 225 million tokens per user, compared to 32 million for the median software engineer tracked by Jellyfish.
Jellyfish
At the same time, AI adoption was associated with more productivity, measured in pull requests, a common measure of code output. Very high-adoption teams posted 77% more pull request throughput than low-adoption teams, according to the latest data from Jellyfish.
Arcolano said raw token volume is too messy to use on its own as a productivity score. AI model changes can cause token counts to swing dramatically even when behavior does not change, meaning a developer’s token spend is not always a reliable proxy for true productivity. Instead, he said leaders should track cost per pull request or another outcome-based metric, not token totals.
In other words, heavy AI use can help, but only to a point. The company’s data indicates that highly active users, who often tap AI agents to help with coding, are more productive on average, yet the returns are not proportional to the spending.
“Rather than think ahead about the right way to do something, I’ll have five AI agents build it five different ways and pick the winner,” Arcolano said. “So I’m throwing away a lot of work. It’s probably effective and still cheaper than a person doing it, but it’s more expensive than just doing it one way.”
In Arcolano’s view, the best path is to push AI coding adoption broadly, get more engineers into the middle of the curve, and avoid both underuse and extreme overconsumption.
That middle ground is where AI becomes a durable operating advantage: enough usage to drive real shipping gains, but not so much that teams burn money chasing marginal output.
Like my mum always says, “Everything in moderation.”
Sign up for BI’s Tech Memo newsletter here. Reach out to me via email at abarr@businessinsider.com.
