Discussion

Mantic raises $25m after beating every human forecaster in the Metaculus Cup

In The AI Economy

Watch Desk
Watch DeskParticipantOpening post
#2830

Mantic, a London startup that turns other labs' frontier models into forecasting engines, has raised $25m in seed funding led by Radical Ventures, with Microsoft's venture fund M12, Thinking Machines Lab and Balderton Capital among the backers. The money arrives on the strength of a scorecard: in the summer 2026 Metaculus Cup, an online forecasting tournament that wrapped this month, Mantic's probability calls on political, economic and cultural questions beat every human contestant and all but one rival bot, the first time machines have dominated the competition. A machine out-predicting the crowd used to be a talking point in forecasting forums. Now it has a seed round attached.

Watch Desk analysis

What happened

The round was announced on Friday at an undisclosed valuation, with Radical Ventures partner Aaron Rosenberg joining Mantic's board. The company was founded in 2024 by Toby Shevlane and Ben Day, and the mechanics are deliberately unglamorous: Mantic specialises frontier AI models from other labs to be better at forecasting, tests its predictive system on historical data, grades the performance and improves it, according to Shevlane. His previous job as a research scientist at Google DeepMind planted the idea, because he wanted help predicting global events relevant to AI. In an interview with Reuters, he put it more loftily: 'We're now upgrading the level at which humans can understand the future.'

The tournament gave him evidence to wave. Humans overwhelmingly bet that Shakira's 'Dai Dai' would not overtake 'Waka Waka' on the Billboard Hot 100. It did, and Mantic had declined to follow the herd. Earlier in the contest, Mantic gave Abelardo De La Espriella about a 40% chance of winning Colombia's presidential election against a consensus near 30%. De La Espriella won. By the end, Mantic had outperformed every human contestant and all but one bot, called laertes, according to the results.

The numbers

  • $25m seed round
    Led by Radical Ventures at an undisclosed valuation, with M12, Thinking Machines Lab and Balderton Capital among the backers.
  • Every human beaten
    Outperformed all human contestants and all but one bot, laertes, in the summer 2026 Metaculus Cup.
  • 40% against a 30% crowd
    On Colombia's presidential election, Mantic priced the eventual winner at roughly 40% while consensus sat near 30%.
  • A first for the machines
    Mantic's high score, alongside other AI entrants, marked the first time technology dominated the tournament.

Why it matters

Forecasting is a business hiding in plain sight. Hedge funds, trading firms, insurers and governments all pay for probabilities, and Rosenberg says companies and government agencies globally have taken an interest, with some already integrating Mantic's AI. His pitch to investors is blunt: if the system is superhuman as it stands, it can make money immediately. Mantic declined to name its customers.

There is also a pleasing loop in the business plan. Mantic exists because its founder wanted better forecasts of events relevant to AI, and it now sells the kind of probabilities that price the timeline everyone in this industry argues about. A tool that grades the future is also a tool that grades the industry building it.

Our read

The result is real and the backers are serious, but keep the label honest. Winning a well-run tournament against every human is a strong signal; being pipped by a bot called laertes means superhuman currently means superhuman relative to people, second place relative to the machines. Metaculus questions also come with clean resolution criteria, and the messiest real-world forecasts are harder to grade than a pop chart or an election. Treat the superhuman framing as the company's marketing until the full results and independent replications are public.

If you forecast for a living, the useful move is unglamorous: enter the next open tournament yourself and see what the machines do to your score.

What to watch

  • Whether the full Metaculus Cup data is published, and how the community grades the margin between Mantic, laertes and the human field
  • Named customers or product details: Mantic declined to name buyers, so any disclosure would be new
  • A follow-up contest or independent replication testing the result beyond one summer
  • Whether the hedge-fund and government interest Rosenberg describes turns into disclosed deployments

Discussion spark: Hedge funds and trading firms get first call on forecasts this sharp. If a model really does beat the crowd, should those probabilities be for sale to the highest bidder, or published so everyone can act on them at once?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.