Ethan Mollick says Anthropic's newest AI is almost as good as human experts

Ethan Mollick, a professor who writes the One Useful Thing newsletter, says Anthropic's new AI model, Claude Fable 5.2, is "almost" artificial general intelligence. That term means an AI as good as most human experts at most thinking tasks done on a computer. He calls the current state "Jagged AGI": expert-level in places, well behind in others.

What Mollick tested

Mollick used Fable 5.2 before its public launch. It follows Anthropic's earlier model, Opus 5.5. He asked it to build the most impressive launch demo it could. It produced a video in one try, with a twist ending. He says it showed creativity, judgment and playfulness, at the level of a human expert making a movie with code alone.

He also pointed it at a tiny USB keyboard called a Codex Micro. With no public instructions, the AI read the code of the two apps that control it. It then worked out how to light the keys and read the knob and joystick. It wrote six games, including Simon and a safecracking game played with the knob.

The business test

Mollick then told Fable to launch a business that would earn him at least $20,000 in a year, using no more than 20 minutes of his time a month. The AI screened 22 ideas. It chose Prebid Brief, a service telling equipment sales reps which park and water projects are being funded or designed before they go out to bid.

To produce its first issues, it read public agendas from 201 city and county governments. Mollick teaches entrepreneurship. He says he would have been impressed if a team of MBA students had handed in this plan at the end of a semester.

Where it falls short

The AI's own estimate was a 20% to 40% chance the business would succeed. Mollick stopped before it began sending public-records requests to local governments. He worried it would take money and fail to deliver quality. Some of the setup it assumed would be easy was not in place, and he does not know how it would handle sales and marketing.

His view is that an AI running a business on its own is likely not here yet. With a few hours of human work a week, he thinks a real attempt could be made. The headline claim of the piece is that the AI still lacks self-directed judgment and taste. Its mistakes are no longer simple errors or made-up facts, which makes the gaps harder to spot.

What is not known

This is one person's account, and the tasks were chosen by him. The material does not say when Fable 5.2 launches publicly or what it will cost. It also does not show whether the AI could actually run a business for a full year.

What to do

• Treat AI plans and prototypes as strong first drafts that a person reviews.
• Be careful before letting any AI spend money or contact customers for you.
• Ask the AI to rate its own odds, as Mollick did, then check its reasoning.