AI Coding - what is it good for ?

When I am out and about I hear a lot that AI coding is only useful for small projects or scripts, but the Omarchy Linux distribution is a useful counter example to this.

David Heinemeier Hansson (DHH) is perhaps best known as the creator of Ruby on Rails, and he is also the creator/author of Omarchy.

DHH says the Omarchy 4 branch is now around 30,000 lines of new code, and that the majority of this was was written by GPT 5.5.

It's worth noting that Omarchy is not an example repo, it's a public Linux distribution project with thousands of commits, substantial community attention, and a mix of Shell, Lua, QML, CSS, Python, and templates.

This does not mean the AI built it alone, DHH’s makes the point is that you still need review, taste, architecture, and judgment (Fusion Coding in action).

I was interested in what the token economics may be for this, and a rough estimate is that 30,000 lines of code might represent perhaps 300K-600K output tokens (depending on language and formatting).

At GPT 5.5 API pricing of $30 per million output tokens, the generated code cost would only be around $9-$18.

Of course, Agentic Coding burns far more than final code i.e. prompts, repo context, file reads, diffs, reviews, retries, tests, and dead ends etc. A more realistic interpretation may therefore be be 5M-50M input tokens plus 0.5M-2M output tokens, which at GPT-5.5’s listed $5/M input and $30/M output would land roughly in the $40-$310 range before caching discounts.

With heavy repeated context or lots of failed attempts, it could be higher, but with prompt caching and tight workflows, maybe a little lower, so this seems a happy medium.

A 30K codebase might only represent hundreds of thousands of final output tokens but agentic coding can burn an astonishing amounts of context - a case in point is OpenClaw reportedly consumed 603 billion tokens across 7.6 million requests in 30 days, with a bill around $1.3M, using about 100 Codex instances !)

This is not happening only in public side projects, Google has said that about 75% of its new code is now AI generated and then reviewed/approved by engineers (up from earlier reported figures of 25% in 2024 and around 50% in 2025).

Microsoft’s Satya Nadella has said that roughly 20-30% of code inside Microsoft repos is now written by software, while Meta has talked about AI doing perhaps half of Llama development over a similar horizon.

The point is not that AI makes code free and easy (it doesn't), it is that the bottleneck was never just the cost of attempting a large implementation, it was the coding quality and accuracy and ability to actually do it, and clearly that is getting better with every model release.