DeepSeek's Cheap AI Now Beats Big Rivals on a Tough New Test

Artificial Analysis runs what is essentially a league table for AI models. It scores them on a set of tests and publishes a single number, the Intelligence Index. Last week the site added a brand-new test — a private one, meaning outsiders can't see the actual questions or answers — and retired an older test called τ³. Almost immediately, the top of the table shuffled. A model called Astra had been racking up points and had pulled level with another top performer, Fable. Then a much less famous model, DeepSeek V4.1 Flash, quietly moved into first place.

The awkward part: the site changed how it scored things twice in three days, and critics online say it looked like the rules were being adjusted so a favored model wouldn't seem to fall behind a rival. Whether or not that's true, it's a useful reminder that leaderboards are snapshots, not settled facts. When a test is private, nobody outside the company can check whether the questions are fair, whether they've been seen before, or whether they actually measure anything you'd care about — like writing well, planning a trip, or fixing code.

The bigger signal is who's on top. DeepSeek is the Chinese lab that stunned markets in early 2025 by showing that strong AI doesn't have to cost a fortune to build. If a budget model can top a demanding new test, it pressures every other company's pricing. That's the same pattern we've seen in TVs, phones, and cloud storage: once a cheap option is good enough, the expensive option has to justify itself.

For everyday users, there's nothing to do. But expect two things. First, more capable AI features arriving in the apps you already use at prices that feel almost free. Second, more marketing that says "number one on the leaderboard." Treat that the way you'd treat "award-winning" on a cereal box — real, but narrow.