Kilo Code DESTROYS Fable 5?

**Julian Goldie** (0:00)
So, how do you get Fable 5 level intelligence without Fable 5 itself? Obviously, Fable 5 has been taken down, nobody can use it. But what can you do in the meantime whilst you wait for it to come back up? Well, Kilo Code actually created a recent case study like you see here, and I've actually implemented the model, and I'll show you what we found today, where basically, if you have a combination of different models planning, then you'll get better outputs, and you can get better intelligence side by side. So let's test it out and see. And the thing that I would say with all of this, we saw this with Fusion as well that came out from OpenRouter recently, is like you can try it and just see what you think. I wouldn't pay too much attention to benchmarks. I would just test out yourself. I'll show you what we tested so far. And so the context here is like Fable 5 was the most powerful AI ever made public, and then it was shut down worldwide three days after launch. Now, Kilo, according to their research, tried to prove something, which is if you hand a cheaper model, a good plan, it can build the exact same thing that Fable 5 would, identical, but it requires less tokens or it requires less powerful models to actually get the most out of it.
Again, this isn't my theory. This is the Kilo team that ran the test and published it. So they had two frontier models plan the same service. Then they made both build the winning plan from identical starts. And this is the key point, is all in the planning according to this. So Fable 5, the most powerful model ever made public, 80% on SW Bench Pro and a cheap model produced services that were identical, down to which individual uses 35% rollout enabled, both passing 15 checks. The only difference here was the amount of tokens and the cost. So according to the Kilo blog, planning with Claude Fable 5 and implementing Gpt 5.5 produced the same service for 59% less than using Claude Fable for both phases. So you can get a great plan and then plug it into your system. According to this, so you get like the identical output, but 62% less to build. Across a full plan and build pipeline, that's 59% cheaper, which is a 2.4x gap that scales to roughly, you know, as you can imagine, quite a lot.
So the plan quality, where does the intelligence actually live? And you can see the details here. So they've actually created this blog on, you don't have to use Fable and Mifos to work on the frontier. So basically, what you can have is one plan in and then you get identical builds out, and this is what Kilo are trying to prove. So they gave the same plan to two different models, both passed all 15 acceptance checks, and then the two services were identical.
So you've got the planner, which is like, that's where you use a frontier model, the plan decides all, and then the agents actually build. So you can use cheaper agents for building to get identical outputs.
Now, why does this work? Well, basically, the model stops mattering as much when you have a more pinned plan. So a vague plan would leave forks open. So each model would guess, two models build two different apps, a pinned plan decides every fork, and so there's nothing left to guess with a pinned plan. So any model can land on the exact same build, right? Vague plan, any model, the output differs. Pin plan, which decides every fork, any model, same app, which is pretty crazy. Now, it wouldn't be like any model, but you get the point, like you could use something less powerful, something cheaper, and then you can build out better. Now, there's three different moves here. So the smart model would write the plan, you switch the model in a click, and then the cheap model would build it.
And then you can have review agents, like check it for bugs and security and that sort of thing. So you got the plan, the build, review agents, the check for security, logic and that sort of thing, and then it gets shipped. And so if you look at this, you got the old way and the new way, where, for example, if you plan and build with the best models, that would use up more tokens. If you use the new way, your plan and then you can build with a cheaper model and that would be more efficient, 59% less on some of the tests. And also if one model gets taken down, like Fable 5, it's okay. You can route the plan that you saved to model B. And so the plan survives, you just swap the builder and then you don't have to worry about using one model again.