Mantic beat every human forecaster in the 2026 Metaculus Cup. Mantic also finished second overall. Both of those things are true, because the thing that beat it wasn’t a person either — it was another bot. That single detail tells you more about where prediction markets are heading than the $25 million seed round that followed.
Reuters reported on September 18 that the London-based startup closed the funding after the summer 2026 Metaculus Cup wrapped. Mantic’s system assigned probabilities to political, economic and cultural events and came out ahead of the human field. Global companies, government agencies, hedge funds and trading firms have all come sniffing. Trading firms in particular, which is the part worth sitting with for a minute.
What “superhuman” actually means here
I review AI tools for a living, and the word “superhuman” gets thrown around so casually that it’s basically decoration at this point. In this case it has a specific, narrow, verifiable meaning: in one tournament, over one summer, against one pool of human volunteers, this system scored better. That’s a real result. It is not the same as “Mantic knows the future.”
Forecasting tournaments reward a particular skill set. You need calibration — when you say 70%, things should happen about 70% of the time. You need to update fast when news breaks. You need to resist the human urge to tell a satisfying story instead of quoting a boring number. Machines are structurally good at all three. They don’t get emotionally attached to a prediction they made in June. They don’t get bored reading the fourteenth article about the same election. They don’t round 3% up to “basically impossible.”
So the headline result is less shocking than it sounds. What’s genuinely interesting is that the ceiling in that tournament wasn’t set by a human at all. Mantic didn’t lose to a superforecaster with twenty years of practice. It lost to code. The human leaderboard and the machine leaderboard have separated into different conversations, and the interesting fight is now happening entirely on one side of that split.
The hedge fund tell
Every AI product pitch comes with a list of interested parties, and most of that list is noise. Enterprise pilots go nowhere. Government agencies sign exploratory agreements and forget about them. But hedge funds and trading firms are a different kind of signal, because they are the least sentimental buyers in existence. They don’t care about your benchmark chart or your demo video. They care whether the output makes money, and they will find out within a quarter.
If trading desks are circling a forecasting system, they’ve done their own math on whether the probabilities are worth anything. That’s a harsher review than anything I could write.
It also creates an obvious tension. A forecasting engine that’s genuinely good is worth more as a private edge than as a public product. The customers with the deepest pockets want exclusivity, and exclusivity is the opposite of a broadly useful tool. I’d watch closely whether Mantic ends up as a platform anyone can use or as a very expensive pipe feeding a handful of funds.
The questions I’d want answered
One tournament is a data point, not a track record. Before I’d recommend anyone build a workflow on top of this, I want to know a few things:
- Does the performance hold across multiple tournaments, or was the summer 2026 question set unusually friendly to a machine approach?
- How does it behave on genuinely novel events, where there’s no historical base rate and no thick news coverage ?
- What happens when its own forecasts start moving markets and prediction platforms, and it begins reading the consequences of its own output as evidence?
- Is it calibrated at the tails? Most of the value in forecasting sits in the 1% and 99% range, and that’s exactly where confident systems tend to fall apart.
None of these are gotchas. They’re the standard questions you ask any model that performed well on a benchmark, and forecasting tournaments are benchmarks with better PR.
My honest read
$25 million in seed money for a team that can point to a real, public, adversarial result is not crazy. The result was earned in the open, against opponents who were trying, with scoring rules nobody could quietly reframe after the fact. That’s more rigor than most AI startups can claim about anything they’ve shipped.
What I’d push back on is the framing. “Superhuman forecasting” makes it sound like a solved category. The actual story is that automated forecasters now compete mainly with each other, human volunteers have been pushed off the top of the board, and the people most eager to buy the output are the ones who’d rather nobody else had it. That’s a narrower and stranger development than the headline suggests, and considerably more interesting.
🕒 Published: