Has AI Coding Hit a Plateau? My Numbers Say Maybe.

Photo by Jackson Sophat on Unsplash

I’ve been tracking my AI coding tool usage since October. Not because I’m a data scientist — I’m not. I’m a developer who got tired of relying on vibes to evaluate whether these tools were actually helping.

Every week, I log: how many times I used the AI, how many times I accepted its output, how many times I rejected it, and why.

The data tells a story that nobody in the AI hype cycle wants to hear.

The Numbers

In October 2025, I accepted AI-generated code 72% of the time. By April 2026, that number was 61%.

A 15% drop in six months.

First-attempt correctness — code that worked without modifications on the first try — dropped from 58% to 43%.

Context understanding — the AI correctly grasping what I wanted — dropped from 67% to 49%.

These aren’t small fluctuations. They’re trends. And they’ve been consistent for half a year.

The Obvious Question

Am I getting pickier? Maybe. But I don’t think that’s the whole story.

I’ve been coding for 11 years. My standards haven’t changed much in the last 6 months. What has changed is the tools I’m using and the models behind them.

Three Theories

I have theories about why the numbers are declining. None of them are provable with my data alone, but they’re consistent with what I’m seeing.

Theory one: speed over quality. The models are being optimized for faster inference. Faster inference means shallower reasoning. The AI generates code quicker, but the code is less thoughtful.

I noticed this first in December. Response times dropped from about 4 seconds to about 2 seconds. Around the same time, I started seeing more code that was syntactically correct but semantically wrong.

Theory two: training data pollution. As AI-generated code proliferates on GitHub and Stack Overflow, models are increasingly trained on their own outputs. It’s a feedback loop that reinforces mediocre patterns.

I can’t prove this, but I’ve noticed the AI generating code that looks like other AI-generated code. Same structure, same patterns, same blind spots.

Theory three: feature creep. AI coding tools are adding features faster than they’re improving core code generation. New features get marketing budgets. Quality improvements don’t.

My primary tool has had three major updates in 6 months. All three added new features. None visibly improved the base code generation quality.

The Counterargument

“But the benchmarks are getting better!”

They are. And benchmarks are useful for comparing models under controlled conditions. But my data isn’t from benchmarks. It’s from real work — messy codebases, ambiguous requirements, tight deadlines.

The gap between benchmark performance and real-world performance is growing. Models excel at LeetCode problems and struggle with production code. That’s not a model problem — it’s a benchmark problem.

What I’m Doing About It

I haven’t given up on AI coding tools. They still save me time. They still catch bugs I miss. They’re still worth using.

But I’ve changed how I use them.

I track everything now. Not just acceptance rates, but the specific types of tasks where the AI excels and where it fails. This helps me know when to trust it and when to code manually.

I use multiple tools. Claude Code for complex reasoning, Aider for speed, Copilot for autocomplete. Each has different strengths and weaknesses. Using all three means I’m never dependent on one tool’s blind spots.

I maintain my own skills. I still write code without AI regularly. Not because I’m a purist, but because I don’t want to atrophy. The AI is a tool, not a crutch.

I give detailed feedback. When tools let me rate responses, I actually do it. Specific feedback, not just thumbs up/down. If we don’t tell them the quality is declining, they won’t know.

The Uncomfortable Part

I think we’re in a hype cycle. AI coding tools are genuinely useful — I’m not disputing that. But the narrative that they’re getting exponentially better every month doesn’t match my experience.

They got dramatically better from 2024 to mid-2025. Since then, the improvements have been incremental at best. And in some areas, they’ve regressed.

This isn’t a complaint. It’s an observation. And I think it’s important for developers to have realistic expectations about what these tools can and can’t do.

The Silver Lining

The declining quality has actually made me a better developer.

When the AI’s output was reliably good, I stopped thinking critically about the code it generated. I’d accept it, run the tests, and move on.

Now that the output is less reliable, I review everything carefully. I catch more issues. I understand my codebase better. I make more deliberate decisions about what to accept and what to rewrite.

The AI’s decline forced me to be more engaged. That’s not nothing.

Where This Goes

I don’t know. Maybe the trends reverse. Maybe a new model architecture fixes the quality issues. Maybe the tools get better at understanding context.

Or maybe this is the new normal — AI tools that are useful but imperfect, requiring constant human oversight and judgment.

Either way, I’ll keep tracking. Data beats vibes. And right now, my data says: use the tools, but trust your own judgment more.

AI Tools, Coding, Developer Tools, Programming, Software Development, Technology