The state of AI coding in 2025: Adoption, proficiency, and transformation

The different software areas were fairly well represented with most respondents coming from backend, frontend, and machine learning roles.

Different sized companies were fairly evenly represented in the responses.

The respondents were mostly senior software developers with over 65% having 4+ years of professional experience.

AI coding tool adoption has become universal with nearly 98% of respondents using AI software tools several times a week. When broken down by seniority level, the more experienced developers with 2+ years of experience are the ones that aren’t daily active users. In some cases, up to 5% of them only used these tools once a week. Usage patterns are quite consistent when broken down by company size. Note that while several of the plots for junior developers indicate 100% adoption, this may indicate sampling bias.

AI was used to generate more than 50% or more of codebases in the last 30 days by nearly 40% of developers.

Tab complete usage patterns are nearly universal likely due to this being one of the first standout AI features in IDEs. Additionally there is very strong adoption with read-only “how does this work” workflows, debugging, test generation, document generation, UI development, and code review. Security use cases adoption still trails significantly.

Claude Code and Cursor adoption is universally the highest. Github copilot is also well-represented especially among developers from larger companies, likely due to Microsoft enterprise contracts. Stand-alone code review platforms still haven’t captured as much of a developer market share.

Most developers are taking steps to control their agent behavior with the majority adopting some sort of agent-specific configuration file (AGENTS.md, CLAUDE.md, etc).

AI coding platforms are seeing increasing amounts of spend with nearly 25% of developers spending $100+/month. More senior developers are also the biggest spenders with 16% of developers with 10+ years of tenure claiming to spend $1000+/month. Additionally the highest spending tends to happen by developers at either very small companies or very big companies.

Developers are mostly satisfied with their AI coding tool setup. More senior developers are most happy with their setup though satisfaction drops at larger companies, likely due to red tape around flexibility to experiment with and access to different tools.

For most software functional areas, the productivity gains sit normally distributed with a median around 1-2x. Frontend efficiency gains are significantly higher with many developers seeing 2x or more as a multiplier. This is likely due to how prevalent frontend code is in public repositories on the intent. Machine learning and data science work is even peakier in gains around the 1-2x multiplier. For design, mobile, and infrastructure, the normal distribution is much more flat indicating the gains are not as clear across developers.

Across different aspects of the development workflow, the efficiency gains are typically 1-2x or more with especially prominent gains on speed of development. AI software tools also tend to improve how quickly developers are able to learn and pick up new concepts. A concerning observation is that nearly 20% of developers say AI software is making them worse at context switching. This may be due to the additional time needed to supervise AI tasks and how it’s impeding developer’s ability to get into flow state when doing deep work.

While AI software tool adoption has been very strong, developers still report high failure rates due to hallucination, incomplete solutions, and generally not understanding the context of a problem.

The most cited categories of desired improvements are around software agents adopting to developer and codebase practices, better codebase understanding, and reduced hallucinations. A growing percentage also wants even more autonomy from their agents.