Notes from the factory

Opus 5.5 is the engineer I was. Fable 5.2 is the one I am now.

How I compare Claude Opus 5.5 and Claude Fable 5.2 after a few days of heavy use. Opus lets nothing slip and spins. Fable is smarter, misses small things, fixes them, and ships.

Two models get the same feature. On the left, an island labelled Opus 5.5: a person paces in front of a bench with the piece of work on it, while sheets of a list land around them one at a time until the island is covered. On the right, a line of three islands labelled dev, staging and production, with an agent labelled Fable 5.2. Fable pushes the piece to dev; the check there goes ochre and an exclamation mark appears over it, because the sorting is wrong. The agent works on it, the check passes, and the piece goes on through staging to production, where the windows of the town light up. At the end the piece is live, and the person on the Opus island is still pacing among the sheets, with a question mark over their head.

We have been using Claude Fable 5.2 heavily for a few days now, and we were high on Claude Opus 5.5 for a long while before that. I have been trying to pin down when to use which model, what each one is good at, and how they work together. The clearest way I have found to explain it is two versions of myself.

The engineer with the list

About seven years into my career, I was a pretty good engineer. I was responsible for a lot of servers, infrastructure and production deployments, and I kept a list of everything that needed to be done. The list was so complete that I put it on a big screen outside my office door, where anybody could see at any time how many things I had to do.

It wasn’t a good look, but nothing ever slipped through the cracks. I was responsible for a lot of production infrastructure, and I never had an outage and never had downtime.

I also lost a lot of time on things that weren’t that important. At the time each one seemed like the most important thing, because somebody had asked me to do it, or because it was what we were supposed to do. Could we have shipped faster? Could we have shipped more? I’m not sure.

The engineer I am now

After everything I have been through since, I can tell what a feature I put forward is really worth, and where its risks are and where they aren’t. If I submit a pull request, you can be pretty sure it provides the value somebody asked for. There might be a small bug, but there will never be anything catastrophic, and if there is a small bug I’ll fix it. No big deal.

I’m shipping more now than I ever have, and it’s working well. I worked the same way before agentic coding: same mindset, same process. Solve the problem, provide value, move forward, and if something doesn’t work, fix it and keep moving.

Two models, two versions of me

That is how I see Opus 5.5 and Fable 5.2. Opus is the engineer with the list. Fable is the one I am now.

Fable is much smarter than Opus. Sometimes something slips through the cracks with Fable. Nothing major has ever slipped through, but there have been a number of times when we had to do some rework.

That’s okay, because of what we have built around our software factory. A lot of it is traditional, and a big part of it is CI/CD: every change goes to a development environment first, and then up to the environments that matter more.

A sorting bug on the way to production

Last week we were working on database performance for a client. Everything worked locally, but once the change reached Google Cloud there was an issue. It was a minor one, in sorting. The database on our machines and the one in Google Cloud compared text by different rules (its collation), so the same query returned rows in a different order.

Fable pushed the code and ran its tests in the development environment, and found that the sorting wasn’t working as expected.

What’s nice about Fable is that it didn’t stop to tell me it had found an issue and ask whether I would like it fixed. Of course I want it fixed. I want it fixed right in the first place. Fable took the extra effort: it fixed the issue, and it checked its work again as the change went to the development environment, then staging, then production.

Now everything’s fine. The sorting is fixed, and the performance work went out with it. We cut a lot of the client’s database overhead, and their users are getting results much faster, because we were able to ship.

Opus would have spun and spun on all the possibilities and permutations of that feature. Sure, Fable missed one thing. At the end of the day, the feature is in production.

The factory is what makes the trade work

Fable’s misses are cheap for us because of what surrounds it. A change goes through checks, a development environment and staging before it reaches production, and the model can see what broke and fix it on the way. That is where a small miss gets caught, before a client’s users ever see it.

So before we pick a model, we look at how much of the factory stands between it and production. How we build that factory is on the factory page. If you want one around your own agents, get in touch.

Tell us what should exist.

Bring us something you want built, or something painful you want gone. Talk to our agent now, or email us and an engineer answers.

A voice call with our AI agent.

Email us