Ronacher says Fable 5.2 is "crazy good" at code that paints
Anthropic never shipped an image model, and Ronacher finds that funny given what the coding model does with a canvas.
Armin Ronacher says Fable 5.2 is "crazy good" at writing code that paints, and he notes the oddity that Anthropic never shipped an image model of its own.
It's kinda funny how Anthropic didn't do an image model but then Fable 5.2 is crazy good at writing code that paints.
He followed up with a run to back it up. He fed in a single reference image, and everything after it in the post is what the model painted from that reference, in code. His verdict on the output was short.
What he actually showed
One reference image in, a set of generated pieces out, produced by a model whose job is writing programs. Ronacher calls it "getting scary good". That is one engineer's run posted to X, not a benchmark, and he did not say what the prompt was, how many attempts it took, or what language or drawing library the model used. Nobody outside his terminal knows that yet.
The first was the reference image, the rest is what Fable paints from it. This is getting scary good.
For working engineers
The distinction worth holding onto is that this is not an image model. A diffusion model hands you pixels. A coding model handed a canvas hands you a program that produces pixels, which means the result is source you can read, diff, parametrize and re-run at a different size. When it is wrong, you fix a line rather than reroll a seed and hope. That is a different workflow with a different failure mode, and it is why a result like this lands differently than another image generator demo.
It also lines up with how Fable 5.2 has been behaving elsewhere. The model took the top slot on Agent Arena at 1.6x the next model, as we covered earlier. Generative art in code is a reasonable thing for a strong code model to be unexpectedly good at, since the task is mostly a long chain of precise geometry and color decisions held together across a lot of lines, which is the same muscle as any other large single-file generation job.
What Ronacher did not claim is that this replaces anything. He said it is funny, and he said it is scary good. Those are impressions from someone who ships, not a comparison against an image model on any shared task, and there is no such comparison in what he posted.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.
