Mantic Builds an AI Forecasting Engine for Real Decisions

Mantic is a London AI research and product company trying to make uncertainty more usable. Founded in 2024 by CEO Toby Shevlane and CTO Ben Day, the startup builds a generalist forecasting engine for questions that resist a tidy spreadsheet: elections, regulation, market shocks, product demand, geopolitics, and other events where context and judgment matter.

The timing is not subtle. Executives have more data than ever and still get blindsided by the future. Mantic is betting that large language models, structured research, specialized scaffolding, and probabilistic reasoning can turn that mess into forecasts people can monitor and test. Its strongest public proof came in the Summer 2026 Metaculus Cup, where Mantic finished ahead of all 676 human entrants. One other AI system finished higher, a useful reminder that a tournament result is evidence of progress, not a declaration of omniscience.

About Mantic

Mantic focuses on medium-term judgmental forecasting, generally one week to one year out. These are questions where historical data alone may not be enough because the answer depends on policy, human behavior, shifting incentives, or events with few clean precedents.

The product combines horizon scanning, probabilistic forecasts, explanations with references, and ongoing monitoring. Users can track questions that matter to their organization and update views as new evidence arrives. Mantic also offers MCP and API access, which matters because useful forecasting cannot live forever in a separate dashboard. It has to meet analysts, operators, and decision systems where work already happens.

The company presents the product to executives, traders, investors, researchers, strategists, risk teams, and policymakers. Its official funding announcement says hedge funds are using Mantic to assess market-moving events across geopolitics, macroeconomics, and business. It also says Fortune 500 customers are applying forecasts to decisions around M&A and product releases. Those customer classes are company-reported, and Mantic has not named the firms or disclosed contract values.

Why Mantic Matters Right Now

Forecasting has always suffered from an awkward operating problem. Organizations can commission a report, build a quantitative model, or ask a room of smart people what they think. Each method helps, but few can continuously research hundreds of questions, assign probabilities, show the reasoning, and update the answer as the world moves.

Mantic is trying to make that process computational. Its system is designed to work across domains rather than optimize for one narrow market. The company says it is especially useful where pure statistical extrapolation breaks down, such as geopolitical conflict, regulatory change, central-bank decisions, competitive moves, and supply-chain shocks.

That breadth is both the opportunity and the risk. A generalist forecaster can become infrastructure for decision-making, but only if its probabilities stay calibrated across topics and over time. A confident answer without transparent sourcing would merely industrialize bad judgment. Mantic's product emphasis on references, monitoring, and explicit probabilities is therefore not interface polish. It is part of the trust model.

The Tournament Record and the Research Loop

Public tournaments give Mantic something enterprise AI products rarely offer: a visible scorecard against independent forecasters. Mantic's site reports that its system placed 4th among 539 humans in the Summer 2025 Metaculus Cup. In Summer 2026, it beat all 676 human participants in the competition. The published leaderboard still placed the separate bot laertes above Mantic, so the clean conclusion is narrower and more interesting than the headline: AI forecasting systems have moved into the top tier of a demanding human field.

Mantic also created Crucible, an adversarial tournament that rewards people for writing questions on which leading AI forecasters disagree. The point is diagnostic. If one system finds the relevant niche dataset and another misses it, or if models interpret a regime shift differently, the disagreement reveals where the forecasting stack needs work.

That research loop appears in the team's academic work. A June 2026 paper by Matthew Aitchison, Scott Jeen, Shevlane, and Day found that strong AI forecasting ensembles benefit from combining models with complementary errors, not merely taking more samples from the same model. In plain English, a room full of similarly trained machines can develop the same blind spot. Diversity matters when the crowd is silicon too.

Leadership Built Around Technical Judgment

Shevlane came to Mantic after 2.5 years at Google DeepMind, where the company says he worked as a senior research scientist and co-led experiments on Gemini's dual-use capabilities. His Oxford doctoral work examined how organizations govern LLM releases and misuse risk. That background fits a product whose output may influence high-stakes choices: accuracy matters, but so do boundaries, evaluation, and knowing when a system should not be trusted.

Day previously led research at Foresight Data Machines, applying AI to steel-production optimization. He holds a Cambridge PhD in machine learning, with research in meta-learning and graph neural networks, and worked as a research consultant on AI for drug development. The shared thread is not simply model building. It is applying machine learning where the world is noisy, the data is incomplete, and the cost of a plausible mistake is real.

Mantic says its broader team brings experience from Google, Citadel, Palantir, Oxford, Cambridge, Goldman Sachs, McKinsey, and the Bank of England. LinkedIn listed 21 discoverable employee profiles when reviewed on September 23, 2026, although that is a platform count rather than an audited headcount.

Funding, Hiring, and the Scaling Test

Mantic announced a $25M seed round on September 18, 2026, led by Radical Ventures. Balderton Capital, Thinking Machines, DRW, FT Ventures, M12, Episode 1, Charlie Songhurst, and Thomas Wolf also participated. Reuters reported that the valuation was not disclosed. Combined with Mantic's earlier $4M pre-seed, the company has announced $29M in funding.

The company plans to spend the new capital on its team, compute, data, and customer work. It says it is hiring across AI research, engineering, product development, sales, and operations. No dedicated careers page was verified; Mantic directs candidates and partners through its official site and contact@mantic.com.

The hiring mix is the strategic signal. Research and engineering improve the forecast engine. Product work makes probabilities legible inside real decisions. Sales and operations test whether strong benchmarks can survive customer data, deadlines, compliance, and executive scrutiny. Scaling all three at once is harder than winning another leaderboard.

What Mantic Has to Prove Next

Mantic has crossed an important threshold: its forecasting system can compete with expert humans on a public tournament and attract enterprise users and serious capital. The next proof will be less theatrical. It will happen when a trader, strategist, or government team changes a consequential decision because Mantic identified a risk early, explained its reasoning, updated fast, and stayed calibrated.

That is the bigger market shift. AI is moving from generating content about the present to assigning probabilities to the future. The winners will not be the systems that sound most certain. They will be the ones that expose uncertainty clearly enough for people to act without pretending it disappeared.