The launch video Fable cut for the project — using the very scenes it rendered, all from a single vague prompt. Roam the live build at ode-to-yosemite.vercel.app.
When vibe coding, you start with an end objective. This is often a vague idea in your head, not fully thought through, both in terms of the final shape of the output and the many steps to getting there. Apart from just writing code, it's the model's job to: (1) infer the goal as best it can (often by asking you clarifying questions), and (2) make many sensible choices on the path to fulfilling it.
Fable feels like a game-changer (and has that "big model smell") because it excels at these critical tasks in a way no other model does.
My main test started with a broad ask: "build me an accurate, navigable, to-scale 3D rendering of the main attractions of Yosemite Valley that I can roam around and explore. Feel free to ask me clarifying questions."
For models thus far, this would've been too vague a task. Not for Fable. The clarifying questions it asked weren't "who are you building this for" or "what tech stack should I use". Instead, they carried agency ("I'm going to pull satellite images from Esri and elevation data from NASA, that sounds good?") and ambition ("Do you want to be able to walk or fly or both?").
The collaboration felt like working with a super smart and resourceful coworker who seeks your confirmation (if at all) rather than specific guidance.
After a few minutes, Fable ended up building a more than decent v1. It combined the satellite images with the elevation data and turned Yosemite into a 3D space I could navigate with my keyboard in the browser.
But something felt off. There were no trees! The instinctive solution to a problem like this would be to randomly fill the space with procedurally generated trees—and I would have been okay with this!
But Fable, as it often does, went a step further. Its solution was to first analyze the pixels on the satellite images, classify the ones that could be trees, and add them only in those spots. In the process, it realized it could also detect meadows, rivers, and snow, and proactively decided to add those too!
If my initial ask was vague, the final output far exceeded anything I could've imagined. I was astounded and delighted. And Fable continued making me feel this way with every follow-up task:
- When asked to add weather conditions, it made the snow build up over time and the fog fade away as the day passed. When asked to add lighting, its night scene included the Milky Way.
- When I asked it to upload the code to GitHub, it realized that the intention of the project was to showcase its own capabilities, so it tailored the README to do exactly that.
- For v1, I did a screen recording. For v2 I asked Fable to use the scenes it created to build itself a cinematic launch video. It did so with flying colors.
As I worked with it, I realized that my own ambitions for the project were growing, and Fable was meeting me every single step of the way. It felt magical in a way that is hard to describe.
Apart from this task, I also used Fable for some brainstorming and thinking through life decisions. This is always murky territory, but I was impressed by its ability to connect seemingly disparate thoughts and ask me sharp and insightful questions. I want to return with more curated context and spend more time with the model here.
This is a controversial model release, and many of Anthropic's decisions around its pricing and usage are extremely concerning. More on that later. I also haven't tested it for day-to-day software development, the kind that matters for serious projects.
But I think it is worth taking a step back and appreciating just how much of a step change this model is, and rethinking the ambitions of whatever you're building or doing accordingly.