model-routing

Claude's new Opus 5 model is out and on most benchmarks it's actually beating Fable. I'm going to talk about why no one saw this coming, when you should use Fable versus Opus, and whether there's still room for models like ChatGPT Sol or open-weight models like Kimi K3.

So Fable 5 has been out for about a month now and it's truly amazing until you run out of usage on the subscription and have to pay the API costs. Opus 5, on the other hand, is a slightly smaller model.

And so there's really no reason to think that it would be better than Fable 5. Across most of the benchmarks, it absolutely is.

It also beats ChatGPT Sol across almost every benchmark on both price and performance. Before this release, ChatGPT could at least claim they were the cheaper version of Fable.

Now, not so much. I've been using Opus non-stop for the last three days and it's hard for me to see why anyone would choose ChatGPT 5.6 Sol right now.

But that doesn't mean that Opus is the only model you should be using. Fable is still the best model in a lot of important ways...

of its work. So here's how you use them together.

If you're doing anything simple or straightforward, just use Opus. It's going to end up being cheaper, and it's going to do a great job.

But if you're doing anything complex and important—you're doing something where you need the model to have good taste and judgment— You need it to reason through many different factors and priorities. Then you should use Fable on high effort with one caveat.

When you're giving Fable a task, include the extra instruction that you're in charge of design and review, but all implementation should be handled by Opus sub-agents. Ultimately, you want Fable to come up with the plan and you want Fable to be given the final sign-off.

But Opus is actually better at the implementation. And you'll also end up saving a lot of usage this way.

Currently, on Claude subscriptions, you can only use up to 50% of your usage on Fable. I've been working this way for the last three days and only about 30% of my usage is going to Fable.

Fable will typically spin up three to six sub-agents that all run Opus...

ption, you're only paying 4% of the usage cost, which is about 8% of Kimi K3 right now. And while Fable is only available on the Max plan, Opus is available on the Pro plan.

So unless you're an enterprise that's forced to pay for usage, I think Claude's where it's at right now.

Anthropic just released Opus 4.8 and it has had a mixed response. So let's talk about where it excels and where you might want to continue using Opus 4.6.

4.8 was built on top of 4.7, and that's where a lot of the criticism seems to stem from. If you don't know, 4.7 also had a mixed response.

Many people, our agency included, have continued using 4.6 for most work. That's because 4.7 has been overly sterile, takes everything literally instead of applying common sense judgment, and tends to use a lot more tokens.

Opus 4.8 is getting a lot of those same criticisms, but it's also really impressive in some areas. Anthropic only released the benchmarks where it's performing very well, but it does perform well on the benchmarks.

Notably, it's the most honest and least lazy model currently. They specifically designed it to be effective on long running tasks.

And so far, that seems like where it really shines. If you have a very large or difficult project that you just want it to work on and you don't mind spending a lot of tokens, 4.8 might be the best model available right now...

d from the desktop app. But hopefully Anthropic notice the response from the community and decide to keep it around for now.

4.8 is also not going to be the smartest model for very long. In that same post, Anthropic announced that a new class of model will be coming out in the coming weeks.

And of course we know this to be the methods class. And of course we can expect methods to be incredibly good at coding.

And also incredibly expensive. It seems like Anthropic is currently all in on creating the most powerful models for software engineering.

And it makes sense. That's where all their money is coming from.

And it lets them use the models internally to develop their own product. It seems like they may be leaving the door wide open for ChatGPT to come in and make Codex the go to product for knowledge work.

For now, I think it's just important to remain agile. Make sure your data is structured in a way that you can easily swap out the model as needed.