Anthropic's largest study of real-world coding-agent usage found that the more skilled you already are, the more an AI assistant multiplies your output — a pattern the researchers call persistent returns to expertise. The data, drawn from roughly four hundred thousand sessions across about two hundred thirty-five thousand people over half a year, shows that AI coding tools amplify existing expertise rather than flatten the gap between novice and master.
Key facts
- What: By studying hundreds of thousands of real coding sessions, Anthropic found that experienced engineers get more out of AI assistants, not less, a direct challenge to the idea that AI levels the playing field.
- When: 2026-06-25
- Primary source: read the source
The scale and source of the data make the finding credible. Anthropic analyzed roughly four hundred thousand coding sessions from about two hundred thirty-five thousand people, gathered from late 2025 into spring 2026. This is not a survey of opinions or a handful of lab volunteers — it is a record of real engineers doing real work with the tool, analyzed in a privacy-preserving way so the company studies patterns across the crowd rather than reading any one person's project. It is, in effect, the largest look anyone has published at how coding agents get used in the wild.
The reason experts pull ahead comes down to what an AI coding agent actually is. It is not a vending machine that spits out finished software when you press a button. It is more like an extremely fast, tireless junior engineer who needs direction. You have to describe the goal precisely, break a big task into the right pieces, notice when the output is subtly wrong, and steer it back on course. Every one of those is a skill, and they are exactly the skills that experience builds. A seasoned engineer knows what to ask for, can smell a bad answer, and can catch the kind of mistake that compiles cleanly but breaks in production. A beginner, handed the same powerful assistant, may not yet know enough to tell good work from plausible-looking garbage, so they get less leverage from it, not more.
The dynamic is like a power tool. Hand a nail gun to a master carpenter and they frame a house in a fraction of the time. Hand the same nail gun to someone who has never built anything and the speed does not help much, because the bottleneck was never how fast they could drive nails. It was knowing where the walls go. AI coding agents move the bottleneck from typing to judgment, and judgment is precisely what expertise is.
The result cuts against a popular hope and a popular fear at the same time. The hope was that AI would democratize software, letting anyone build. The fear was that AI would make experienced engineers redundant. Anthropic's data suggests both are too simple. Instead of replacing experts, the tools appear to be amplifying them, which has real consequences for hiring, training, and how teams decide where to put their best people. It connects to a thread running through this whole year of AI news, from the finding that AI now writes most of Anthropic's own code to the cautionary tale of a company that burned through its yearly coding budget in four months because powerful agents are powerful spenders too. For more on what these autonomous helpers are, see our explainer on AI agents.
The honest caveat sits at the center of the study: Anthropic is a company studying how people use Anthropic's own product, and "expertise" and "returns" are slippery things to measure from usage logs. The company built the analysis carefully and shared its methods, but a self-interested party measuring its own tool always deserves a second look, ideally an independent one. It is also worth remembering what the finding does not say. "Experts gain more" is not the same as "beginners gain nothing," and the long-run picture — what happens as today's beginners grow into tomorrow's experts using these tools the whole way up — is exactly the part no six-month snapshot can answer yet.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)