London AI startup Mantic raises 25 million dollar seed after beating humans in forecasting contest

Mantic, a London-based artificial intelligence forecasting startup, said on September 18, 2026 that it raised $25 million in seed funding led by Radical Ventures. The round also included Microsoft's venture fund M12, the startup Thinking Machines Lab, Balderton Capital and other investors. The financing comes after Mantic's system outperformed every human competitor in the summer 2026 Metaculus Cup. Chief executive and co-founder Toby Shevlane, a former research scientist at Google DeepMind, said the result shows machines are getting better at predicting uncertain real-world events. 'We're now upgrading the level at which humans can understand the future,' Shevlane said in an interview.

Mantic was founded in 2024 by Shevlane and Ben Day, who serves as chief technology officer. The startup focuses on judgmental forecasting, which deals with messy questions involving business, technology, geopolitics, policy and culture, often weeks or months into the future. Unlike weather prediction or conventional statistical models that rely on clean datasets, Mantic's system is built for situations where research, context and judgment matter as much as historical data. The company says it makes medium-term predictions, from one week to one year out, and benchmarks its system against human forecasters.

The summer 2026 Metaculus Cup was an online forecasting tournament in which participants assigned probabilities to political, economic and cultural developments. The competition wrapped up this month, and results showed that Mantic had assigned probabilities more accurately than human competitors. Its high score, along with those of other AI entrants, marked a first in which technology dominated the competition. WebProNews reported on September 17 that an AI system won the Summer 2026 Metaculus Cup, with two other AIs grabbing second and fifth place, while humans took third and fourth. The Economist reported on September 17 that the result marked the first machine victory in a seasonal contest.

Mantic does not build its own frontier foundation model from scratch. Instead, the startup specializes frontier AI models from other labs to be better at forecasting. It tests its predictive system on historical data, grades its performance and then improves it. Shevlane said his previous work at Alphabet's Google DeepMind sparked the idea because he needed help predicting global events relevant to AI. In the Metaculus Cup, one advantage Mantic had over human forecasters was avoiding a herd mentality, he said. In one question, humans overwhelmingly bet that pop star Shakira's 'Dai Dai' would not overtake 'Waka Waka' on the Billboard Hot 100. That turned out to be wrong, but Mantic did not bet heavily on the consensus.

Key Facts

Reuters reported on September 18 that Mantic raised $25 million in seed funding led by Radical Ventures at an undisclosed valuation. The round included Microsoft's venture fund M12, the startup Thinking Machines Lab, Balderton Capital and others. Aaron Rosenberg, a partner at Radical Ventures, has joined Mantic's board. Tech Startups reported on September 18 that Sky News first reported the financing, with Reuters later confirming the $25 million round and additional details.

The Metaculus Cup results, which resolved on September 5, showed Mantic outperformed every human contestant and all but one bot, called laertes. The broader showing from AI systems marked the first time machines dominated the competition, according to Reuters. Mantic's system also made specific calls that diverged from human consensus. Early in the tournament, it gave Abelardo De La Espriella about a 40% chance of winning Colombia's presidential election, compared with a roughly 30% consensus forecast. De La Espriella ultimately won.

Mantic says it makes medium-term predictions from one week to one year out about geopolitics, business, policy, technology and culture. It has an edge on topics where a purely data-driven approach is infeasible or insufficient, according to its website. Its product capabilities include horizon scanning, forecast generation with reasoning and references, and monitoring that updates predictions as new information emerges.

The team includes Shevlane, who spent 2.5 years at Google DeepMind as a senior research scientist and co-led a team within the Gemini effort that designed experiments to test Gemini's dual-use capabilities. Day, the CTO, was previously head of research at Foresight Data Machines, developing AI for optimizing steel production, and holds a PhD in machine learning from the University of Cambridge. Mantic says its team has experience from Google, Citadel, Palantir, Oxford, Cambridge, Goldman Sachs, McKinsey and the Bank of England.

Context from the Forecasting Research Institute in July showed several models statistically indistinguishable from superforecaster accuracy on its tournament leaderboard. Cassi AI led the pack there, with systems from xAI and Google DeepMind following close behind. On market-style questions, one Cassi entry even surpassed the superforecaster median for the first time. But earlier this year, teams of Metaculus Pro Forecasters still held a narrow edge. Spring 2026 head-to-head tests across 99 shared questions gave the human group a 1.25-point advantage per question on average.

Analysis

What this really means is that AI has crossed a threshold in a domain long considered resistant to automation. Judgmental forecasting requires synthesizing incomplete information, weighing competing narratives and updating beliefs as events unfold. Mantic's system did not just beat average participants; it outperformed every human contestant in a competitive tournament. That result, combined with $25 million in seed funding from a syndicate that includes Microsoft's M12 and Thinking Machines Lab, suggests investors believe forecasting is a viable commercial application of AI. The company's strategy of specializing frontier models from other labs, rather than building its own foundation model, is a pragmatic template for vertical AI startups.

The commercial interest is already visible. Aaron Rosenberg of Radical Ventures said hedge funds and trading firms were particularly eager for Mantic's forecasts. 'If Mantic is superhuman as it is, they can make money off of that immediately,' he said. Rosenberg also said companies and government agencies globally have shown interest, with some already integrating Mantic's AI. Mantic declined to name its customers, which makes the traction harder to verify, but the interest from hedge funds is a strong signal of perceived value.

The bigger picture here is that forecasting could become a new front in the AI race, alongside chatbots, coding agents and image generation. Mantic is not alone. The Forecasting Research Institute data from July showed Cassi AI leading a pack of models statistically indistinguishable from superforecasters, with xAI and Google DeepMind close behind. In the Metaculus Cup, one bot called laertes actually finished ahead of Mantic. So Mantic's edge is real but not absolute. The Spring 2026 head-to-head tests, in which Metaculus Pro Forecasters held a 1.25-point advantage per question across 99 shared questions, further show that humans remain competitive in some settings.

However, the trajectory is clear. AI systems are improving quickly, and the Metaculus Cup result is the first time machines have dominated a seasonal competition. Mantic's ability to avoid herd mentality, as shown by the Shakira question, highlights a structural advantage: AI does not feel social pressure to conform to a consensus. That said, AI can also have correlated errors if trained on similar data or if it overfits to historical patterns. The most likely future is not total replacement of human forecasters but a hybrid approach, where AI handles large-scale probability estimation and humans provide context, judgment and accountability.

Why It Matters

For businesses and governments, accurate medium-term forecasts can improve strategic planning, risk management and investment decisions. Mantic targets corporate executives, traders, investors, researchers, strategists, regulation and risk teams, and policymakers. If AI can consistently beat human experts and superforecasters, it could change how organizations allocate resources and prepare for uncertain events. The company's product includes reasoning and references, which may help users trust the predictions, though trust in AI forecasting will need to be earned over time.

For the AI industry, Mantic's success demonstrates that frontier models can be specialized for high-stakes judgment tasks. The fact that Mantic beat all humans but one bot shows that model size alone is not enough; training, scaffolding and data selection matter. This could inspire more startups to focus on vertical AI applications that adapt existing models rather than compete on building the largest foundation model. It also raises questions about accountability: when an AI forecast informs a government policy or a financial trade, who is responsible if the prediction is wrong?

For the future of expertise, the story is nuanced. The Spring 2026 results showing human Pro Forecasters with a narrow edge suggest that human judgment still adds value. But the trend line points toward AI systems becoming more capable. What this means for professional forecasters is not necessarily obsolescence but a shift in roles: humans may become supervisors, data providers or ethical overseers of AI forecasting systems. The bigger picture is that AI is extending human judgment into new domains, and the organizations that learn to combine both will likely have an advantage.

Next Up

Mantic plans to commercialize its forecasting system. Rosenberg said companies and government agencies globally have shown interest, and some have already integrated its AI. Hedge funds and trading firms are particularly eager. The next milestones to watch include customer announcements, product launches and whether Mantic can maintain its forecasting edge in future Metaculus Cups or other benchmarks. The startup will also face competition from Cassi AI, xAI, Google DeepMind and laertes, the bot that beat it in the summer tournament.

Regulatory and ethical questions may also arise. Forecasting political events such as elections, as Mantic did with Colombia, could draw scrutiny if the technology is used to influence outcomes or if its predictions are inaccurate. Mantic's use of public information and its offer to integrate customers' proprietary data sources may raise privacy and fairness concerns. With $25 million in fresh capital, the company has room to hire and refine its technology, but the seed round is only the beginning. A future funding round may reveal a valuation and test whether investors remain confident in the superhuman forecasting thesis.