Vibe coding is entering the mainstream, according to new research from cloud firm Fastly, with senior developers now routinely relying on AI to build and bench test their work.

In a survey of almost 800 devs, Fastly found that almost a third of developers (32%) with more than ten years of experience admit that more than half of their shipped code is created using AI, while around 36% said that up to half of their work is AI-generated.

Surprisingly, junior developers are more likely to do the bulk of the work themselves, despite being arguably more familiar with the vibe coding capabilities of AI. Fastly found that just 13% of those with less experience in the job, up to two years, use AI to build 50% or more of their code.

For those relying more heavily on AI, the benefits are clear. Senior devs reported significant speed gains using AI, with 26% agreeing that it makes their work a lot faster, compared to only 13% of junior developers.

More than half (59%) of senior developers agreed that AI tools have helped them ship faster overall, compared to 49% of juniors. 

But, that speed is being hard won, with senior developers spending more time fixing AI code, and 30% spending so much time editing AI output that they said it negated any efficiency gains.

Despite its growing popularity, vibe coding still appears to be rough around the edges, with 65% of all developers saying they frequently have to go back over AI-generated code, and just 3.5% saying they rarely need to edit AI output.

Even so, for the majority (80%) of devs, AI has made their job easier and more enjoyable, with only 14% reporting it has had no real impact and just 4% saying it has made their work less satisfying.

According to Fastly, this disconnect comes down to simple psychology, with AI coding feeling smoother and initially faster than having to create code from scratch, even if significant time is instead spent on correcting errors.

“An AI coding tool like GitHub Copilot greatly helps my workflow by suggesting code snippets and even entire functions,” one respondent said. “However, it once generated a complex algorithm that seemed correct but contained a subtle bug, leading to several hours of debugging.”

Recommended reading

• The Biggest LLMs Are Generating Vulnerable Code by Default
• How to Avoid Wasting Months on AI That Doesn’t Work
• Anthropic Reveals How Hackers Are Really Using AI

Earlier this year, Anthropic’s Economic Index similarly found evidence of a culture shift toward vibe coding among developers. 

Analysing half a million interactions with its Claude and Claude Code models, the AI firm found that 79% of all conversations were identified as automation, meaning that the AI was directly performing coding tasks, with just 21% labelled as augmentation, where an AI and human work in collaboration to complete a task.

However, while Fastly’s research indicates that developers using AI in this way are faster, other studies suggest otherwise. 

In July, the nonprofit research organisation METR measured the productivity impact of AI on experienced open-source developers and discovered that when using AI, tasks took 19% longer to complete.

METR noted that the gap between expectation and reality was striking, with devs initially believing that AI would speed their workflows by as much as 24%, and even after experiencing the slowdown, still thought that AI had increased their speed by 20%.