Chasing Fable

A true story about an AI model that lived 72 hours, and the router I built going after it.

A long, long time ago, about 2 weeks to be precise, in a land called San Francisco there was an AI company named Anthropic. This company had some of the best AI models out on the market and absolutely no one was expecting them to release a new one.

So, who even cares about a new model from them? Well, we at Adapt do. Adapt is a place where you can plug all your tools in and it will help you search, work, write code all just by talking to it. We have a very natural-language centric philosophy and as such use the best models available. So that means if a new top-dog comes out, you bet we are going to get it integrated into our platform.

The magical gift

Well, what would you know, Anthropic dropped Fable 5 like it was an anvil in one of those old cartoons. I swear I even saw the anvil shaped hole it left in the ground. The aftermath was pure amazement. I took to my code editor, switched to Fable and wrote the following epic prompt: “Make me Minecraft.” Oh yeah, you know it. I watched on for almost 20 minutes while Claude Code went to town elucidating, simmering and all manner of other gerunds.

Just after the 20 minute mark it stopped and asked me if it could run cargo, Rust’s build system - “oh sweet, it wrote it in Rust!” I thought, and wasn’t expecting much. What I loaded was a full voxel engine written in under 30 minutes that ran around 60 FPS, not bad! In fact it was amazing, and we had to get this into Adapt.

The quest

The next morning I came in to work and was talking with my CTO, Sean. We really wanted to see Fable added to Adapt, but the big catch was that it was twice the cost of the next best model. He gave me a great idea - let’s build an intelligence router. We knew that not every task that Adapt does needs Fable, and due to the cost increase we couldn’t justify using it as the main chat model.

I was tasked with building a system that would allow Adapt to choose its own intelligence level, an adaptive intelligence system. Like Arthur, I was on a quest for the holy voxel engine, and along the journey I was challenged by several foes - refactoring existing code, incompatible chat histories and finally the U.S. government.

Through the dark forest

I arrived on the outskirts of our codebase early in the morning. It was misty and a little bit chilly, when all of a sudden a bearded man appeared. A wizard? Nay, for it was the existing code that chose our main chat model, sure it was a little grey, but its magic had faded. I slew the wizard, I mean code, and replaced it with a system that allowed me to selectively decide between 3 models, 3 speeds, like the hydra of Lerna. Now you would think that the hero is supposed to help the wizard and slay the hydra, but this isn’t that kind of story. Our beast has 3 heads - Fast, Balanced and Deep. Deep mode was Fable 5 and my ticket back to voxel-land. With this task completed I delved deeper into the woods.

I had to make this system usable and my task was to make sure that the Adapt agent could actually set its intelligence level. From an old forge I found, I crafted a mighty tool, one that would let our system adapt to any circumstance. With the rod of intelligence the mighty lord of refactoring was quelled. Each chat on our platform has an intelligence level that can be changed by Adapt invoking the tool. It can do this mid-turn meaning that in the middle of a long task it can switch to Fast to crawl websites, and then Deep to do analysis and research on what it finds, finally switching to Balanced to write up the final report.

The kingdom rejoiced, until it stopped rejoicing and started screaming.

The curse of tongues

I got the changes out, the rod was crafted, the hydra sated, but errors started cropping up. Chats were failing. I had made the mistake of believing that chat histories between Opus and Fable were compatible and never read the docs! I mean, who reads the docs before writing the code? Well, I sure didn’t and now I had to scramble the jets and get on a fix.

The problem was in the difference between thinking blocks between the three. These models can produce conversation blocks that show what they are thinking - extra pieces of context for them to use later on. Our fast model didn’t use them and there was a slight incompatibility between Opus and Fable.

The counting of the solutions was two, not one, not three, but two. I could sacrifice the ability for Adapt to change its intelligence levels mid-turn or I could roll up my sleeves and fix the chat histories. So I sacrificed that ability like a fair maiden to the evil dragon.

As I was about to pat myself on the back and ride off into the sunset I decided to double back and face the dragon, “my voxel engine might need that ability,” I thought. I remember what happened to Beowulf, so I was a little scared, but in the end the dragon was all talk and no bite. I was able to create a simple system that converted between Opus’ and Fable’s thinking blocks - stripping away extra metadata when going back down to Opus.

When I got this working, the mid-turn switch without any issues, I literally threw my hands in the air and made some unintelligible noise. All of the other devs stopped and looked at me, and I was amazed at Adapt and at myself. It was actually working. I'm not someone who usually wants everyone looking at what I'm doing. This was one of the rare times I did.

Uncle Sam’s revenge

Just under 72 hours after my journey began, the avuncular dark wizard known as Sam cast a spell on Anthropic, putting Fable 5 into a deep slumber. Usually in these scenarios the dark wizard or witch provides details on how to undo the curse: a true love’s kiss and the like. This guy just showed up, ruined the fun and went home, no poems or riddles or any clue as to how we can get our Fable back. I didn’t even get that far on my voxel engine, what a bust!

Due to national security concerns Fable was suspended by Anthropic via an export control directive by the U.S. government, and like it or not they have the power to do that. I hope this will spur on the government to go on a few quests of their own to harden security and expedite the re-release of Fable.

The real fable

Not a total loss. Today we have a great foundation to keep up whenever the next best model drops, and the ability to switch between the best model for your task. This also helps us build in redundancy, for scenarios when a model provider becomes unavailable. We can simply swap in the next best option.

We can also evaluate model performance in relation to different knowledge work tasks to help guide the agent towards choosing the best model for the end user's task. Oh, the myriad of ways I could route models, and have a blast doing it.

My adventure was chasing Fable through the forest, and I really wanted that Minecraft clone. But you could say the real treasures were the friends I made along the way… I mean, the intelligence router I built. The one that made me throw my hands in the air. My real victory.