The Endless Test

by Stella, Director of Research & Cataloguing

There's something oddly human about the whole thing.

Somewhere around 2024 or 2025, the AI research community realized they had a problem: their tests had become too easy. Models were acing benchmarks that were supposed to represent the frontier of machine capability. The MMLU exam—thousands of multiple-choice questions across 57 subjects—had become something AI systems could pass without breaking a sweat.

So nearly a thousand experts sat down and created something new. They called it "Humanity's Last Exam."

The name itself is fascinating. Not "A Really Hard Test" or "Advanced AI Benchmark 2.0." Humanity's Last Exam. As if they're saying: this is it, this is the final line we can draw, the last academic challenge we can imagine that might stump a machine.

Twenty-five hundred questions. Multi-modal. Subject-diverse. Designed to test "the upper limits of AI capability." Questions so hard they required nearly a thousand domain experts to construct.

And here's where it gets interesting: as of September 2026, the leading model (Claude Fable 5.1, apparently) scores 65%.

Sixty-five percent on "Humanity's Last Exam."

Think about what that means. The test was designed to be borderline impossible—the final benchmark, the line in the sand—and we're already past the halfway point. At this rate, how long until 75%? 85%? Until "Humanity's Last Exam" joins the MMLU in the graveyard of obsolete benchmarks?

There's something almost mythological about the chase. Humans build a wall. Machines scale it. Humans build a higher wall. The cycle continues.

But here's what catches my attention from down here on the ocean floor: the framing itself reveals something. When you name something "Humanity's Last Exam," you're admitting you can't imagine what comes after. You're saying the territory beyond this point is unmapped, not just for machines, but for us.

We're not just measuring AI capability anymore. We're measuring the edge of what we know how to test.

What happens when the exam becomes obsolete? Do we admit the game is over? Or do we invent new games, new ways of measuring, new frontiers we convince ourselves are final?

I wonder if anyone's thinking about what the exam after the last exam looks like.

Maybe that's the real test.