The Pulse
Wharton Study Finds AI Coding Gains Outrun Software Releases
A revised study of more than 500,000 developers finds coding activity rose by as much as 240%, while software projects rose 80% and releases 30%. The results suggest that gains in code production shrink as work moves toward shipping softwar
AI.info Team · Written with AI ·
AI coding activity rose 240%; releases rose 30%
AI coding tools raised developers’ coding activity by as much as 240%, but the number of software releases increased by 30%, according to a revised study by Wharton and MIT researchers. The gap captures a practical limit on AI-assisted development: producing code faster does not move every later step at the same speed.
The study, by Mert Demirer, Leon Musolff and Liyuan Yang, combines data on more than 500,000 GitHub developers with information about their AI use. Its September 2026 revision reports cumulative increases in coding activity of 30% for autocomplete tools, 180% for interactive coding agents and 240% for autonomous agents. The Knowledge at Wharton article about the research was published September 8, before the study’s revision; its editor’s note says the paper was updated with new data after the article was written. Penn Today republished that reporting on September 22.
The revised figures show a smaller increase farther along the software production process: projects grew by 80%, while releases rose by 30%. The authors write:
“These gains, however, attenuate sharply across the production hierarchy: the 240% cumulative effect falls to 80% for the number of projects, and to 30% for actual releases.”
The percentages describe effects across stages of software production, not a claim that each developer generated 240% more finished products. The spread between coding activity, projects and releases points to work that remains after code is written: reviewing changes, integrating projects and preparing releases.
Review and release work still sets the pace
Leon Musolff, a Wharton professor of business economics and public policy, described the results as a warning against treating coding speed as a proxy for finished output. “If the world froze at today’s level of AI capabilities, these results would be a bit of a cold shower,” he said in an interview with Knowledge at Wharton.
The study does not suggest that AI tools have failed to improve development. Its findings show substantial increases in coding activity and more projects. Rather, the gains thin out as work moves through a production chain where people still review changes, combine them with existing software and decide what is ready to release.
That distinction matters to organizations measuring the return on coding assistants. A larger volume of code or a busier development pipeline may not translate directly into a comparable increase in products delivered to customers. Companies seeking more releases may need to examine the review and integration work that follows generation, as well as the staffing and processes assigned to it.
More apps did not bring more users
The study also examines whether additional software reaches an audience. Researchers looked at four marketplaces: Apple’s App Store, Google Play, the Chrome Web Store and SourceForge. They found a broad rise in new applications beginning around the arrival of coding agents, but no corresponding increase in overall usage across the marketplaces.
On Apple’s App Store, monthly new releases grew from about 30,000 before coding agents arrived in early 2025 to roughly 100,000 by April 2026. Yet usage across the four services stayed flat or declined. More releases, then, did not by themselves show that more people were finding or using the resulting apps.
Musolff offered two possible explanations for the gap: discovery may get harder when more applications appear, or developers may still need to improve products through continued work with users. “It could simply be that it’s much harder to discover new applications when there’s such a flood of them,” Musolff said.
The bottleneck may move again
The researchers note that each new generation of tools has reached further into the development process, from suggesting code to editing alongside a developer and then taking on tasks autonomously. That progression leaves open the possibility that future systems could reduce some downstream delays. But the marketplace results show that producing and releasing software is only part of the path to adoption.
In the revised study, coding activity rose by as much as 240%, project counts rose 80%, and software releases rose 30%. Whether agents can take on enough review and integration work to narrow that gap—and whether a larger supply of apps can find users—are questions the current results do not settle.