SILICON VALLEY — The AI industry descended into crisis Tuesday after Madison LLM outperformed OpenAI’s Astro and Anthropic’s Fable in every major benchmark, including mathematics, coding, and knowing when a meeting should have been an email.
Astro scored 94% on advanced reasoning. Fable scored 97%. Madison scored 100%, corrected two questions, and emailed the examiners: “Let’s make this harder next time.”
In coding, Astro built a website in thirty seconds. Fable built one in twenty. Madison launched a profitable company, sold it, and returned before either competitor had finished explaining how to install the dependencies.
Its context window was described as “everything you’ve ever said, including that thing in March you thought we’d moved past.”
Efficiency proved equally alarming. While rival models required enormous data centers, Madison ran on one iced coffee and a bag of Sour Patch Kids. Researchers briefly achieved a further performance boost by telling it someone else thought they were smarter.
Sam Altman reportedly requested access to the model. Madison placed him on a waitlist.
Dario Amodei called for urgent safety testing. Madison agreed and asked whether he would feel safer holding someone’s hand.
Both companies have since announced ambitious roadmaps.
Madison has announced lunch.