GoKawiilResearch laboratories are weighing up the costs and benefits of paying for artificial-intelligence tools.Credit: J Studios/Getty
James Zou has spent “well over US$100,000” on artificial intelligence in the past year. “These models are very useful for researchers, for coding, for analysis, for literature summaries,” says Zou, a biomedical-data scientist who leads the AI for Science Laboratory at Stanford University in California.
In his view, the fees, which are in the same ballpark as the cost of supporting a postdoctoral fellow at Stanford, are worth it. He says “we’re entering into a new golden age of science with AI assistance” that enables fundamental scientific advances because of the “increasing capabilities of these AI scientist agents”.
But AI assistance is starting to look more expensive for researchers. AI providers have struggled to make the economics work for them on subscription plans, so are hiking up prices and tightening usage limits. In January 2025, Sam Altman, chief executive of the California-based company OpenAI, posted on social-media site X that the firm was losing money on its $200-a-month ChatGPT Pro subscriptions because people were using the chatbot more than the company expected, driving up OpenAI’s use of computing power and electricity.
GitHub, a platform that allows developers to store and share their code, is the latest provider to change its pricing policy. On 27 April, it announced that it would move GitHub Copilot, its AI tool that helps users to write code, from a subscription-based service to usage-based billing from 1 June, citing the higher demands of agentic AI.
Cost–benefit analysis
Those changes are having an impact on researchers who have become increasingly dependent on AI. Attila Gáspár, an economist at Central European University in Vienna, has been using AI to extract data from historical documents. For around 18 months, Gáspár did everything he wanted with his university-paid subscription to the AI chatbot Claude, but in late April, he encountered a problem. “It said, ‘You have hit your limit,’” he explains.
AI is saving time and money in research — but at what cost?
Matteo Niccoli, a geoscientist who uses AI for technical research, upgraded from a Claude Pro to a Max subscription and still hits limits on heavy workdays. Serious scientific projects, he says, involve “multi-session” work, with repeated back and forth between coding, reasoning and analysis. And he’s finding that even those subscriptions “don’t cut it”, meaning he has to work by hand — taking more time and limiting the possibilities of big-data analysis.
The decision of whether AI is worth paying for isn’t only about price and usage limits. The errors AI introduces into its workflows can add more work to a researcher’s load that outweigh the technology’s benefits. Niccoli describes the bottleneck he encounters now is the “thinking and the discussion” around the work: checking the model’s outputs, noticing when its context has become overloaded and knowing when its answers are beginning to drift. “It’s all on you to figure out how to reliably use them,” he says. That makes AI useful — but not necessarily a labour-saver.