Opus 5 Is Just Fable 5 But Cheaper, In Both Directions

Anthropic's announcement art for Claude Opus 5, from their launch post.

I've been building with Opus 5 for a while now, including a full session rebuilding this site.

Here's my one-line verdict. Opus 5 is Fable 5 with a smaller bill, and the discount shows up in two places: one you'll like, one you won't.

The Cheap That's Good

On most coding tasks I throw at it, Opus 5 lands in roughly the same place Fable 5 does.

It writes production code. It holds a plan across a long session. It follows the conventions in a codebase instead of inventing its own.

If you're willing to watch it work and check the output, you're getting near-flagship results for a fraction of what the top tier costs. That's not a small thing when you're running these models daily on your own dime.

The Cheap That Costs You

The discount also buys you a model that's lazier about the stuff it wasn't explicitly told to do.

Here's the example from this week, and it's a good one because it cost me real hours.

I opened a fresh session to work on this site. Fable 5, in that same spot, would have pulled the latest version of the repo before touching anything, because that's what you do when you start a new conversation about a project that lives on GitHub.

Opus 5 didn't think to. It just started building on whatever was sitting in my local checkout.

That checkout turned out to be 19 commits behind the live branch. So it spent hours writing a design system fix that I'd already shipped three weeks earlier, under the same token name, with different values.

It gets better. Somewhere in there it told me, confidently, with file paths and line numbers, that a color I could see on my own site wasn't in my CSS at all. It was in my CSS. Just not in the copy it was reading.

That's the failure mode in a sentence. It wasn't wrong about the code it had. It was wrong about which code was real, and it never thought to check.

To Be Fair, That's a Prompt Problem

I want to be honest about this, because it would be easy to turn one bad session into a verdict.

Almost everything I just described is fixable with better prompt and loop engineering. "Pull the latest from origin before you start" is one line in a project file. Any decent setup should already have it.

So this isn't really a capability gap. It's an initiative gap, and initiative is exactly the kind of thing you'd expect to get thinner as the price comes down.

But it's a change worth naming. If you're moving from Fable 5 to Opus 5 and you've been relying on the model to fill in the obvious unstated steps, you're going to feel it before you can explain it.

Who Should Actually Run It

If you review what the model produces, Opus 5 is an easy call. You get most of the quality and you keep most of the money.

If you've been treating your coding model as something you fire off and come back to, be more careful. The gaps aren't in the code it writes. They're in the steps it doesn't think to take, and those are the expensive ones.

The tell is that every problem I hit was a context problem, not a reasoning problem. Give it the right starting state and it does excellent work. Let it assume the starting state and it will happily do excellent work on the wrong thing.

Watch it, and it's the best value in the lineup right now. Don't, and you'll pay the difference back in rework.