SAN FRANCISCO — Helion Labs unveiled its long-teased reasoning model Thursday morning, an event the company had spent six months calling a leap in machine cognition and which it chose to demonstrate, on the first tip-off day of the NCAA tournament, by filling out a bracket that picks Duke, Kansas, UConn, and Houston in the Final Four.
The model, internally codenamed Athene-3 and externally codenamed Athene 3.0 because someone in marketing thought the period made it sound humbler, was introduced during a livestream in which executives spoke for forty minutes about “agentic reasoning chains” before asking the model who would win it all. Athene-3 said Duke. The CEO smiled the smile of a man who had paid for that answer.
Helion’s launch materials describe Athene-3 as capable of “multi-step deliberation across uncertain domains,” which in practice means it spent eleven seconds generating a bracket statistically indistinguishable from the one printed in the Sunday paper. It picked one 12-over-5 upset, the same one ESPN’s analysts picked, and confidently advanced North Carolina past a team Helion’s training data appears to believe still exists.
A subscription tier offering “premium bracket reasoning” was made available at $40 a month. Asked what distinguishes the premium tier, a Helion product manager said it allowed users to ask follow-up questions, such as why the model picked Duke, to which Athene-3 reliably answers that Duke has “strong fundamentals and a proven coaching tradition,” a sentence that has appeared in every March preview written by a human being since 1987.
“It’s behaving exactly as we’d expect a frontier model to behave,” said an engineer on the post-training team who asked not to be named because the launch embargo had not formally lifted and also because he wished to remain employed. “It read everything ever written about basketball and concluded what everyone who reads everything ever written about basketball concludes, which is Duke.”
The company’s investor deck, which leaked roughly forty minutes after the livestream, projects $2.1 billion in revenue from “decision-support applications,” a category that includes bracket generation, fantasy lineup optimization, and a forthcoming product that will help users decide what to order at restaurants. The deck does not specify what percentage of these decisions will be Duke.
Competing labs reacted with the measured grace the industry is known for. A rival firm pushed out a blog post within the hour claiming its own model, released in January, had also picked Duke, and had done so using forty percent fewer tokens. A third lab announced it would release a model next week that picks Kentucky, which it described as “a differentiated approach to bracket reasoning.”
By Thursday afternoon, Athene-3’s bracket had already lost a 5-seed in the opening window, a result the Helion communications team characterized as “within expected variance for a probabilistic system operating in a high-entropy environment,” a phrase that also describes Helion’s CEO’s father, who picked the same bracket for free.
Helion confirmed the model would be made available to enterprise customers beginning next week, in time, the company noted, for the Sweet Sixteen, by which point most of its picks will no longer be in the tournament.