Fable 5: the model that surpassed our AI usage habits

By Nate (Nate's Substack) · June 23, 2026.

Nate publishes his review of 'Fable 5,' a model he describes as 'the most capable I have ever tested,' despite the fact that it is no longer available: according to his account, the U.S. government pulled it from production days after its launch and Anthropic disconnected it worldwide. The author recorded the review before all that happened, briefly pulled it when the news came, and then decided to publish it anyway, arguing that 'staying silent about what this model can do would be the wrong response to losing access to it.'

The article's central argument is not technical but behavioral. Nate describes the moment he realized something had changed: he had handed Fable 5 a deliberately poisoned database—full of phantom records, corrupted files and planted traps—and, instead of staying to supervise the process as he always does with new models, he went off to do something else. When he came back, the work was finished: not answered, but done. The database was clean, the junk was quarantined (not 'fixed' on the sly), and the model had spontaneously built a review queue with all the cases it was unsure about, as if anticipating that it would be checked.

That autonomous behavior—completing a real task, making decisions, flagging its own doubts without being asked—is what the author identifies as Fable 5's qualitative difference from previous models. And he warns: if used the way AI assistants are usually used (summaries, rewrites, quick drafts, code snippets), that difference will not be noticed. The early reviews calling it 'overkill' are not wrong, he says; they simply accurately describe the smallness of the task being put to it.

The article's main thesis is that 'Fable 5 is the first model that is bigger than our habits.' For three years, the model was the limit: the user learned where it failed and framed their requests below that line. With Fable 5, Nate claims that for the first time it was he who ran out of work he knew how to delegate before the model ran out of capacity. The ceiling is no longer in the model; it is in the user's imagination about what tasks they can hand over.

He calls this 'detailed task imagination': the ability to conceive complete jobs—not prompts—that can be entrusted to a model of this level. He maintains that it is a concrete and learnable skill, but that almost no one teaches it because the last three years have been devoted to teaching prompts instead of jobs.

The article promises to develop the following content blocks for subscribers: why AI has seemed smaller than the headlines (the request-size problem); what 'asking big' means in practice and where Fable 5 still fails; the definition and demonstration of task imagination, with a task carried from start to finish; how to restructure a workweek around jobs instead of prompts; and 'The Whole-Job Spec,' a nine-field template the author fills out before handing any real job to the model.

Nate also explicitly clarifies something that, he says, fills his inbox: no, you cannot rebuild Fable 5 from a system prompt or by stacking smaller models. He claims to have tested those recipes and that they do not work.

The general tone of the article is that of someone who wants to separate the genuine energy he has felt testing the model from the usual fatigue with AI launches. He describes that when he shares these ideas—in his community, in his inbox, with operators—the response he receives is not fatigue but hunger: 'People are not tired of AI. They are tired of being told it is amazing while their actual experience of it remains small.'

As sector context, the narrative that the bottleneck has shifted from the model to the user is a growing debate in the agentic AI community: as models become more capable of executing multi-step tasks autonomously, the differentiating skill ceases to be knowing how to program the model and becomes knowing how to precisely define what work is handed to it.