April 20, 2026
A couple of things: using LLMs for production code absolutely makes you more productive, but if you are working on top of a large, existing codebase, both the scale and shape of that productivity are hard to precisely measure and describe. Existing code wasn’t written with AI (or AI in mind). Big codebases are hard for today’s LLMs (though that’s changing fast).
The point where AI tools crossed over to be consistent productivity wins for most professional programming work was end of November last year. Opus 4.5, November 24. Then GPT 5.2-Codex roughly caught up to Claude on Feb 5th *this year.* We are only 4 months into this new era. Before Opus 4.5 and the Claude Code harness, these tools were mind-blowing but often burned just as much time as they saved (or more). Four months is the blink of an eye. We are all retooling now to figure this stuff out.
But we do have some examples.
Anthropic clearly got here before the rest of us. All in on these tools and co-evolving their models and harness. They are shipping like crazy. It’s truly impressive.
iOS App Store submissions are up 2x year over year. I’m sure all the new apps aren’t great. But this is a big deal. Lots of people (including me!) could do an app now as a side project (or first programming project) that couldn’t/wouldn’t have before. Tagging in @swyx here who has thought about this more then I have.
And speaking of side projects, we shipped a large, ambitious thing last week that I’m really proud of. I’d wanted to build a massively multi player, 100% LLM game since I first played with GPT-4. I started poking at this with Claude Code last October. It got bigger and bigger. Other people started contributing. And now it’s a real thing. And the first thing of its kind: multiple LLM inference loops running all the time for every player, with dynamic UI generation, and a bunch of interesting sub-agent management internals. This project would not exist without coding tools!