Hand-written code is over. For those who can pay

DHH is right. Writing code by hand stopped being the default, and the language that matters now is natural language. I’m in no position to disagree, because it’s been months since I wrote a whole file by hand.

In his Rails World keynote he said 37signals has gone “pencils down” and that hand-written code there is now “an exceptional state”. When the agent gets something wrong, you fix the agent and leave the pencil in the drawer. The new HEY ships as six native apps with a Rust backend. He called Rust “the ugliest programming language” of “probably the last 40 years” and said he loves it, as long as an agent is the one writing it.

A year ago I wrote almost everything by hand. I always followed this technology closely and tried to get as much out of it as I could, and because I knew its limits so well, I didn’t trust it. This new batch changed that. Anthropic’s 5.5 models, GPT-6.1 Sol and GPT-6 Astra are absurdly good. They still get things wrong, but most of the time I now trust them to figure it out. Today pretty much 100% of my code goes through AI. I still review everything and send a lot of it back, and I don’t think I’m good at this yet. I already wrote that the craftsman became the architect. DHH just said the same thing from the stage of the framework he created.

What I’ll concede

The HEY-in-Rust numbers (99% less CPU, 95% less memory, the whole peak load running on a Raspberry Pi) come from DHH himself, about software that hasn’t shipped yet, and he called the Raspberry Pi math “back-of-the-envelope” himself. The 150,000 lines he says he produced in August alone can’t be audited by anyone outside 37signals either.

Even so, I don’t think he’s inflating results to sell something. He told the failure story too: designers built the last Basecamp 5 features with agents, 20 or 30 of those PRs left the architecture looking “like a Swiss cheese”, and Basecamp 5 shipped with a lot of hand-written code. His takeaway was that backing away from agents would be the wrong lesson, and I agree with him there.

He isn’t alone, either. Guillermo Rauch, who runs Vercel and makes a living selling JavaScript, put it bluntly:

Rust wins for the reason I argued before: the compiler becomes the agent’s cop. If nobody needs to find the syntax pretty, the winner is the language that catches the model’s mistakes before they reach production.

Bun is proof, but only partial proof

Bun moved from Zig to Rust with agents and around US$165k in tokens. That proves the agent factory works, but Bun had Anthropic behind it, a huge test suite and access to a model the rest of the world couldn’t even use yet. I can study it and try it on my own projects. I just can’t treat it as a recipe.

The hole: the bill

This is where I stop agreeing. DHH talks from a place where tokens don’t dent the monthly budget, and I live very far from that place.

The common argument right now is that prices dropped. Within three days, Claude Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon all launched at the same price: US$2 per million input tokens and US$10 per million output. The fight just moved somewhere else. Argon was limited to about 650 security professionals, and that price is introductory, later going to US$4/20. OpenAI launched Ultrafast mode, which costs 6 times more to go faster, and a US$500/month Pro plan that includes it. Its always-on agents start at the US$100 plan. List prices tied, and speed and access became the expensive product.

I pay for Grok, OpenAI, Google and Claude, all on the US$20 plans, and the plan is to keep one or two. In reais, that already hurts. We’re broke over here. I use weaker, cheaper models, and that has a cost that never shows up on the invoice. It comes out of my time, in review and rework.

On the other side there’s Theo. He changed T3 Code’s tagline:

He didn’t pay Anthropic ten million dollars. That number is the token value at list price, summed across everyone contributing to the project, and it’s his own estimate (“probably”). The money that actually left anyone’s pocket is a different number. Theo lives on subscriptions. He says a US$200 Claude Code subscription gets you about US$9k of Opus usage per month if you price it at API rates. He says he has “6+ accounts”, that the Max plan is “an absolute steal”, and that API pricing is “insane”. For me the split is simple: the API is for integrations, for when you’re going to sell a product. If the software is yours and only you will use it, a subscription does the job.

He’s right about that, and I have to adjust my argument. Subscriptions are far cheaper than list-price tokens. He calls Claude Code and Codex “subsidized”, even though he also thinks the labs still profit on most of those subscriptions. I had been comparing frontier costs by price per million tokens, and that math is wrong.

The subsidy just doesn’t solve my problem. To get US$9k a month of Opus, you first need to pay US$200 a month. To burn through limits the way Theo does, you need several of those. He celebrated maxing out 4 Claude accounts in a week. I’m not rich enough to have four top-tier accounts of anything. I have four US$20 accounts and I’m thinking about cutting two.

And I’m still on the good side of this math. I can afford two plans. Plenty of people can’t afford one. Even with two, try building a business on US$20 plans alone. The limit runs out midweek, you wait for the reset, switch models, review more. You can learn that way. Competing with someone who has six top-tier accounts is another story. And a Brazilian founder also gets squeezed from the other side, because the market for selling software is shrinking too. I wrote about that in software for one.

Price per token dropped and the top plan is a great deal, but the discount only reaches people who can already pay for the expensive plan, ideally several times over. Volume is still a privilege.

Writing got cheap, maintaining is still expensive

A week later, Theo took heat for saying he’s probably not going to maintain the TypeScript-in-Rust rewrite:

I prefer his honesty, and it shows exactly where “pencils down” stops working: the agent writes an entire TypeScript in Rust, but somebody has to answer for it a year from now, and that person’s salary doesn’t come out of Anthropic’s pocket.

Omarchy itself puts this split in writing. DHH launched a US$100k bug bounty on HackerOne, and under the program rules an AI-generated finding without a manual check and a working proof of concept gets nothing. I don’t see hypocrisy there, because every project that has drowned in AI reports learned the same thing. The agent sweeps, the human signs off, and the bounty money goes to whoever signs.

Primeagen switched sides

Primeagen spent years as the resident AI skeptic. In September he joined Omarchy Core to lead agentic QA with Oligarchy, a harness that boots Omarchy in VMs and lets agents type and click like a person.

If even Primeagen is putting agents on QA for DHH’s distro, I think trust in these models has reached a point few people expected. It went about the same way for me: I kept using everything I could, waiting for the models to get to where I could trust them. When they got there, I went with what fits on my credit card.

DHH is right about the craft, and I put the pencil down too. But as long as frontier models and maintenance are billed in dollars, putting the pencil down the way he did will stay a thing for people who can pay.