AI Coding Tools Hit a Plateau Nobody Wants to Admit
Cursor is worth $29 billion. Windsurf merged with Devin. The actual productivity gains haven't changed much since 2024.
Surveys say 78 percent of developers now use AI coding tools. Cursor's valuation hit $29 billion. Cognition acquired Windsurf to merge it with Devin. Google launched Antigravity. The market is crowded and expensive.
The productivity research tells a different story.
Studies consistently show that AI coding assistants speed up simple, well-defined tasks. Boilerplate. Config files. Test scaffolding. Repetitive refactors. For these, the gains are real. 40-50 percent faster in controlled experiments.
For complex tasks—debugging, architecture decisions, understanding unfamiliar codebases—the gains are minimal or negative. Developers spend time reviewing AI suggestions that look plausible but miss edge cases. They accept code that compiles, passes lint, and still does the wrong thing.
The 2024 Uplevel study found no statistically significant difference in pull request throughput between teams with and without AI assistants. I haven't seen 2026 data that contradicts this at scale.
Cursor boasts 40ms response time for code suggestions. That's impressive engineering. But response time isn't the bottleneck. The bottleneck is that suggestions need to be right, and rightness requires understanding context that extends beyond the current file.
Windsurf's answer is Fast Context and Codemaps—systems that maintain relevant context across thousands of files. In practice, this means the model knows more about your codebase. It doesn't mean the model understands your codebase. The difference matters when you're making changes that have non-obvious implications.
Devin 2.0 claims 83 percent more completed tasks per compute unit compared to its predecessor. The framing is revealing. They're measuring completed junior-level tasks, the exact category where gains were already established. Nobody publishes benchmarks on "correctly refactored legacy systems" or "debugged race conditions in production."
The pricing models are starting to reflect the reality. Cursor's $20/month unlimited access is probably unsustainable as compute costs rise. Devin charges per Agent Compute Unit—$2.25 each on top of a subscription. Pay-as-you-go suggests the vendors know that heavy usage at flat rates loses money.
Where I think the tools actually earn their cost: reducing the activation energy for small tasks.
Before AI assistants, fixing a typo in a file you've never opened required navigating to it, understanding enough context to make the change safely, and testing the result. With a good AI tool, you describe the change and review the diff. If the change is genuinely small, this workflow is better.
The error is extrapolating from "good for small changes" to "good for all development work." The tools are optimized for the former. The marketing implies the latter.
Devin's acquisition of Windsurf signals where the market is heading. The pitch for autonomous coding agents requires combining IDE-level context with agent-level autonomy. Whether that combination actually produces useful software remains unproven at scale.
My current setup: Cursor for daily editing with aggressive review of everything it suggests. Claude Code for terminal-based refactors where I want more control over the process. No agent-driven development for anything I'm responsible for shipping.
The tools keep improving. The benchmarks keep climbing. The gap between demo performance and production value persists. I'll update my workflow when the research shows actual gains on hard problems, not just faster completion of easy ones.