Report: Claude Code Exceeds GitHub Copilot for Merging Pull Requests

An analysis of pull requests published today by Jellyfish, a provider of an engineering intelligence platform, finds that Claude Code is the artificial intelligence (AI) coding tool used most often to merge pull requests (31%), followed by GitHub Copilot at 23%.

Based on analysis of 100 million pull requests made in September by 276,000 developers working at more than 1,300 organizations, the report also notes that AI agents were, on average, involved in 28% of all the merged pull requests tracked by Jellyfish. However, the median for automated merges without human review was just 0.5%.

Overall, the report finds pull request throughput rose by about a factor of two since the analysis was last run in August.

Dr. Nicholas Arcolano, head of research for Jellyfish, said the report suggests that AI coding agents are not replacing the need for software engineers to manage workflows. In fact, the median company ended the third quarter at 98% of its first-quarter headcount while spending 3.8 times as much on AI per engineer. Organizations are discovering there is no substitute for human judgment when it comes to making tradeoffs across the software development lifecycle, noted Arcolano. A human is ultimately held accountable for the quality of the application, he added.

Over the past year, the median company’s cost per merged PR fell by about a third, from $1,202 to $820. Organizations merged more PRs per engineer to reduce those costs while salary costs held about flat. AI’s direct slice of the total cost per PR remains small at about $18.

All that increased AI coding activity, however, is not leading to a corresponding increase in application releases. The rate at which DevOps teams release software has increased 48%. While that represents a significant gain in productivity, it also suggests that DevOps teams are creating more code than ever that is not making it into production environments, noted Arcolano. The issue that DevOps teams need to address next is the processes currently used to build and deploy software using AI, he added.

Finally, the analysis finds most DevOps teams continue to favor proprietary AI models. In fact, the same percentage of engineers chose an open-weight model in September but that percentage is still 3.5 times greater than it was last March. The most widely used open-weight models are Kimi (63%) and GLM (38%).

Just about every organization is now using AI coding tools to varying degrees to generate code. As more AI agents are deployed the challenge is quickly pushing the limits of DevOps teams to manage them. Many DevOps teams are finding they are now spending more time babysitting AI agents at a level of scale that exceeds the ability of human software engineers to track and manage, noted Arcolano.

At the end of the day, the rise of agentic AI engineering is an exercise in change management. The issue is that when it comes to absorbing that level of change no two organizations are going to be equally prepared.