Fable 5.1 vs. Astra: which model wins which job

By Nate.

Nate says he asked two frontier models — Anthropic's Fable 5.1 and Astra, OpenAI's new model — to build the same Mac app from the same short brief, with the initial idea of walking away with a single clear winner. The comparison, however, did not turn out as clean as he expected.

Fable's version taught him things he had not even thought to specify in the original brief: for example, how much space the clipboard should take up in the interface, or how much of an image it made sense to be able to inspect before selecting it. According to Nate, seeing examples of real work made it easier to make those design decisions he had not anticipated.

The turning point came down to speed: Astra delivered its first version of the app faster, which let him start using it and asking for changes while Fable was still generating its proposal. As a result, the app he ended up preferring and using was the one that incorporated more rounds of his own feedback. Nate is honest about the limits of his test: he cannot know how Fable would have responded to those same corrections, because he never put it through that additional test.

From this experience, Nate reflects on an underlying bias in how we compare AI models: the model that is already your favorite gets, over time, more work, more context and more chances to polish an answer. You learn, almost without noticing, to phrase requests in the way that model understands best. When you then test a rival with that same style of prompt, you tend to give it less time and fewer iterations to improve its initial answer, so part of the observed difference may come simply from that accumulated practice with your usual model rather than from any real gap in capability.

In terms of his own use, Nate explains that Astra's arrival has made him reconsider how much work he delegates to Claude. He says he increasingly turns to Astra for drafting, though he still prefers Anthropic's Fable 5.1 when he needs to think through an argument he has not yet resolved. He concludes that a single project can perfectly well justify using both models, each for the part of the job it does best.

The email works as a preview of a longer piece for subscribers, which he describes as covering: why response speed ended up being the deciding factor in which app he kept; what value the app he ultimately did not choose still had; how he divides tasks between the two models (writing, research, spreadsheets, email, code and 3D), with criteria on what to give each one and how to read their answers; and when it makes sense to go for a cheaper option, distinguishing the parts of a long assignment whose instructions are already clear from those that genuinely require human judgment.

As part of that subscriber piece, Nate is also offering two Mac apps that came out of this experiment — named Shelf and Ledge — signed and notarized by Apple and ready to install, along with a guide detailing how he tested both tools, what went wrong with each along the way, and the token and time figures for the comparison. The email does not spell out those specifics, as they are reserved for the full subscriber content.