Consuming too much AI can be a bad idea. New data on 'tokenmaxxing' reveals a better approach.

The tech sector is undergoing a significant shift in how it approaches AI expenditures, as highlighted by new findings from Jellyfish. The research indicates that the most excessive users of AI tokens do not necessarily yield the best results. According to Jellyfish, the top 10% of Claude Code users consumed roughly 10 times more AI tokens than the average developer, yet their output was only about double that of their peers. AI tokens serve as the building blocks for AI models, breaking down words and inputs for processing and providing a pricing structure typically based on millions of tokens utilized. Nicholas Arcolano, head of AI and research at Jellyfish, pointed out that the trend of extreme 'tokenmaxxing'—where users maximize token consumption—is proving to be an unsustainable approach. "CFOs are scrutinizing these practices closely," Arcolano explained in a recent interview. "While companies are eager to invest in AI to accelerate their operations, they also need to demonstrate responsible spending and tangible impacts." Jellyfish’s analysis, based on data from hundreds of companies and thousands of software engineers, reveals that heavy token consumption does not always correlate with increased returns, but rather escalates costs for businesses. This insight signals a shift in the tech landscape toward prioritizing AI efficiency over sheer volume of usage. Arcolano emphasized that even if one believes the value derived from AI surpasses that of human labor, soaring token costs will still raise concerns for financial officers. The report illustrates the stark disparity in token usage among high and low adopters, with top users consuming up to 225 million tokens weekly, compared to 32 million for average engineers. Moreover, while AI adoption has shown a positive link to productivity, as seen through metrics like pull requests, Arcolano cautioned against using raw token counts as a measure of efficiency. He noted that fluctuations in token use can occur due to changes in AI models, making them unreliable indicators of true productivity. Instead, he advocates for focusing on cost per pull request or similar outcome-based metrics. The takeaway from Jellyfish’s findings is clear: while utilizing AI tools can enhance productivity, it is crucial to strike a balance. The data suggests that users who frequently engage with AI for coding tasks tend to be more productive, but the relationship between spending and output is not linear. Arcolano advised adopting AI widely among engineers while maintaining moderation to avoid the pitfalls of both underutilization and excessive consumption. This balanced approach is essential for leveraging AI as a sustainable competitive advantage, ensuring that organizations gain substantial results without overspending.