Ethan Mollick: Anthropic's Claude Fable 5.2 is "almost" AGI, with gaps in taste and self-directed judgment

Ethan Mollick, a Wharton professor who writes the One Useful Thing newsletter, concludes that Anthropic's Claude Fable 5.2 is "almost" AGI, or what he calls "Jagged AGI". In a post published November 9, 2026, he says the model performs at human-expert level on several demanding computer-based tasks but still falls short on taste and self-directed judgment. Mollick says he used the model extensively before its public launch, and describes it as following the "already excellent Opus 5.5". Pricing, context window and benchmark scores for Fable 5.2 are not given in the post.

Mollick's definition of AGI

Mollick uses a deliberately narrow definition: "an AI as good as most human experts at most intellectual tasks you can do with a computer." It excludes physical-world work, interaction with other people, and most actual jobs, where computer work is only one task among many. Under that definition, he says the jaggedness that remains is "no longer about simple errors or hallucinations", which makes it harder to describe.

What he tested

A one-shot launch video. Mollick asked Fable to make the most impressive embeddable demo it could for his newsletter. The result was a code-built video with a twist ending, produced in one shot with no feedback. He says it showed creativity, judgment and playfulness at the level of a human expert making a movie with code alone.

Hardware hacking. He asked Fable to turn a Codex Micro, a small USB keyboard with light-up keys, into a game console. Without public instructions, Mollick says, the model read the code of the two apps that control the keyboard. It worked out how to set key colors and read the knob and joystick, then wrote six games, including Simon and a knob-driven safecracking game. He says it worked on the first try.

Launching a business. Mollick asked for a business that would earn at least $20,000 in a year with no more than 20 minutes a month of his time. Fable screened 22 ideas and chose Prebid Brief, a service telling equipment sales reps which park and water projects are being funded or designed before they go to bid. It read public agendas from 201 city and county governments for its first issues. It also built files on what cities already own, how old it is, and what they paid. Mollick stopped the experiment before launch, wary that the business might take customers' money and fail to deliver quality. He says the infrastructure the AI assumed would be easy to create was not in place. The model's own estimate of success was 20%-40%. Mollick, an entrepreneurship professor, says he would have been impressed by the plan from an MBA team at semester's end.

Rebuilding W. H. Auden's 1941 course. Mollick asked Claude to recreate "Fate and the Individual in European Literature", which Auden taught at the University of Michigan and which had 6,000 pages of reading. The resulting site has a piece for each of the 35 major works on Auden's list, plus the blackboard notes, records played and assignments. The model annotated poems using Auden's published scholarship. Claude also used critic agents and fact-checkers, and published the open-source code on GitHub. Mollick says the work would have taken humans months, and that his own cross-checking and other advanced AIs found no major errors.

Original academic research. He asked Fable to find an unsolved empirical problem in entrepreneurship research, locate the data, preregister hypotheses and methods, write the paper, and have other agents critique it. The result is a paper titled "The One-Dollar Company".

Where the jaggedness shows

The Auden project exposed the gaps. Fable's first version was, in Mollick's words, "accurate and joyless", a fifteen-week course website with tests. His request for something beautiful produced a children's game about ringing bells. When asked why a poem's rhythm or word choice mattered, the model wrote confident readings without considering what Auden himself had said. Each fix took the AI minutes, Mollick says, but it noticed none of the problems on its own. He says it is sometimes hard to tell whether this is a taste difference, an underspecified prompt or a real weakness.

What this means

Mollick's results point to a shift in where the limits sit. In his tests, failures were less about factual errors and more about the model not recognizing that its own output was joyless, off-tone or ungrounded in sources. The business test is the clearest limit: planning and research reached what he calls MBA-team quality, but sustained independent operation was not demonstrated, and he judged it unsafe to try. His practical conclusion is that a human supplying taste, direction and a few hours of weekly gap-filling remains necessary. All of this comes from one researcher's informal testing with early access, not from controlled benchmarks, and the AGI label is his own framing under his own definition.