London startup Mantic announced a $25 million funding round on September 18. The company’s “superhuman forecasting” claim stems from a summer Metaculus tournament in which its system assigned more accurate probabilities to political, economic and cultural events than all 676 human participants.
The result is stronger than a typical language-model benchmark because predictions were made before outcomes were known and were scored on how well the probabilities matched reality. It was not an outright win, however: Mantic beat every human participant but finished behind another bot, called laertes.
Training for uncertainty
Mantic does not claim to have built a general-purpose foundation model from scratch. Chief executive Toby Shevlane says the team takes frontier models from other labs and specializes them for forecasting. The system is tested on historical data, scored for accuracy, then refined.
The task is not to name one future but to assign probabilities to possible outcomes. That makes it possible to distinguish a cautious estimate from a confident but lucky answer. Metaculus participants forecast events across politics, economics and culture; scores are recalculated when the real outcome becomes known.
Shevlane gave Reuters two examples. Mantic disagreed with the majority view that Shakira’s “Dai Dai” would not overtake “Waka Waka” on the Billboard Hot 100; the consensus proved wrong. On another question, the system gave Abelardo de la Espriella roughly a 40% chance of winning Colombia’s presidential election, against about 30% in the aggregate forecast. He ultimately won.
Who backed the $25 million round
Radical Ventures led the round, joined by Microsoft’s M12, Balderton Capital, Thinking Machines Lab, DRW, FT Ventures, Episode 1 Ventures and investor Charlie Songhurst. The parties did not disclose Mantic’s valuation.
Shevlane and Ben Day founded Mantic in 2024. Shevlane previously worked as a researcher at Google DeepMind and says he became interested in better ways to assess world events that could affect AI development.
The company plans to spend the money on research, product development, sales and hiring. Its founder says some large hedge funds already use Mantic for financial-market trading. Reuters also quotes Radical Ventures partner Aaron Rosenberg on demand from companies and government organizations. Mantic has not named customers, so the scale of deployment cannot yet be independently verified.
Why one tournament is not proof
The Metaculus competition is a useful test, but it does not establish that a system forecasts every kind of event better than people. Question selection, scoring rules, available data and forecast horizons all affect the result. Humans can be influenced by group opinion, while a model inherits limits from its training data and base models.
There is also a risk of overfitting. Repeated testing on similar historical tasks may not transfer well to rare crises or events without close precedents. Serious decisions call for repeated independent tournaments, open methodology and comparisons with strong professional forecasting teams — not just a general leaderboard.
The key caveat: Mantic beat all 676 human participants, but it did not take first place among all systems; one bot was more accurate.
Financial markets will be an especially demanding test. An advantage can disappear as soon as competitors find the same signal, while misplaced confidence becomes a direct loss. Until Mantic names customers and publishes results from real deployments, claims about economic value remain statements from the company and its investors.
Bottom line
Mantic offers a practical use of AI beyond chat: rather than one confident answer, it produces several possible outcomes and regularly checks probabilities against reality. Beating 676 people makes the result notable, and the $25 million round gives the startup room to test it with customers.
The next threshold is reproducibility. If Mantic can again beat strong human forecasters and other systems on fresh questions, the case for “superhuman” performance will be stronger. For now, the most accurate description is a strong result in one tournament where another bot still did better.
