$25 million. That’s what a London startup pulled in after its forecasting AI outscored every single human in the 2026 Metaculus Cup and still landed in second place overall. The thing that beat it was another bot.
Mantic announced the seed round on Friday, September 18, with Radical Ventures and Microsoft’s M12 among the investors. Reuters covered it. The plan is to scale the forecasting technology. That’s roughly the full set of confirmed details, and honestly, the sparseness is part of what makes this interesting to me as someone who builds bots for a living.
Why second place is the real headline
Most funding announcements lean on a win. This one has a more unusual shape: the humans got beaten, and then the winner was also a machine. If you’ve spent time in prediction tournaments, you know how much that reframes things. Human superforecasters were, until recently, the benchmark everyone else measured against. Now the top of the leaderboard is a bot-versus-bot conversation, and humans are the field being cleared.
For those of us wiring up agents on the weekend, that shift matters more than the dollar figure. It means the interesting competition has moved. The question is no longer “can a model keep up with a careful human researcher?” It’s “which pipeline design wins?”
What forecasting actually demands from a bot
I want to be clear that I have no inside knowledge of how Mantic built their system. But I’ve built forecasting agents, and I can tell you where they tend to fall apart, which gives you a decent map of what a winning one has to get right.
- Calibration, not confidence. A model that says 90% needs to be right about nine times out of ten. Language models are famously bad at this out of the box. They round toward certainty because that’s what reads well in text. Fixing it usually means post-processing, ensembling, or explicit scoring pressure during evaluation.
- Question decomposition. “Will X happen by March?” is rarely one question. It’s five sub-questions about base rates, current signals, and timeline mechanics. Agents that answer the surface question score badly. Agents that break it apart and recombine do better.
- Retrieval that knows when to stop. Search is cheap to add and hard to tune. Too little and you’re forecasting from stale training data. Too much and you drown the context with low-signal news chatter that pushes the estimate around at random.
- Resolution discipline. Half the errors I’ve logged in my own agents came from misreading what would count as the event happening. Tight reading of resolution criteria is unglamorous and it’s where points hide.
- Aggregation. Running one model once is a coin flip with extra steps. Running many passes and combining them, ideally with different prompts or different models, smooths out the noise.
None of that requires a giant lab. It requires patience and a scoring loop you trust. That’s the encouraging part of this story for hobbyists. Forecasting is one of the few AI tasks with a built-in, unforgiving grading system. You cannot vibe your way to a good Brier score.
What I’d want to know next
The public facts don’t tell us whether the win came from clever architecture, heavy inference spend, a well-tuned ensemble, or some combination. Each answer would imply a different lesson. If it’s mostly compute, this is a story about budgets. If it’s mostly pipeline design, this is a story anyone with API access can learn from.
I also want to see durability. Tournament performance over a single cup is a snapshot. Forecasting skill shows up across many questions and many months, especially when the question mix changes. A system tuned on geopolitics may sag on markets or sports.
The practical takeaway for builders
If you’ve been looking for a project that forces you to write real evaluation code instead of admiring your own prompts, build a forecasting bot. Pick a set of open questions, log your predictions with timestamps, score them properly, and keep a changelog of what each version changed. You’ll learn more about model behavior in a month than you will from a year of chat experiments.
The money going to Mantic is a signal that investors now think this skill is worth productizing, not just benchmarking. Radical Ventures and M12 don’t write seed checks for party tricks. Something in the numbers convinced them the accuracy is transferable to work people pay for.
The humans lost this round. But the tooling that beat them is mostly public components arranged well, and arranging components well is a thing you can practice. That’s the part I find genuinely motivating: the frontier here is accessible in a way most AI frontiers aren’t.
🕒 Published: