Artificial intelligence companies have spent years trying to make machines better at writing, coding, analysing documents and generating images.
A London startup is pursuing a different challenge: predicting what happens next.
Mantic has raised $25 million in seed funding after its artificial intelligence system demonstrated unusually strong forecasting performance in the 2026 Metaculus Cup, an online competition in which participants assign probabilities to future political, economic and cultural events.
The funding round was led by Radical Ventures, with participation from Microsoft's venture fund M12, Thinking Machines Lab, Balderton Capital and other investors. The company's valuation was not disclosed.
For the startup industry, Mantic represents a potentially important shift. AI is increasingly moving beyond generating content and towards systems designed to help businesses make decisions about uncertain future events.
Mantic Is Building AI Specifically for Forecasting
Mantic is not attempting to build a completely new general-purpose foundation model from scratch.
Instead, the company takes advanced AI models developed by other laboratories and specialises them for forecasting.
Its system is tested against historical information where the eventual outcome is already known. Mantic can then evaluate how accurate its predictions were and use those results to improve its forecasting process.
The approach is designed to answer questions differently from a conventional chatbot.
Rather than simply asking, "What will happen?", a forecasting system might estimate that an event has a 30%, 55% or 80% probability of occurring.
That distinction matters because real-world decisions are rarely certain.
Its AI Beat Human Forecasters
Mantic attracted significant attention following the summer 2026 Metaculus Cup, which concluded this month.
The competition tested participants on forecasts covering politics, economics and culture.
Mantic's system ultimately outperformed every human participant and all but one competing bot, according to results cited by Reuters.
The broader results were notable because AI systems dominated the tournament, marking an important milestone for machine forecasting.
That does not mean AI can reliably predict every future event.
But it suggests specialised systems may be becoming increasingly useful at assigning probabilities to uncertain outcomes.
Avoiding the Crowd May Be One Advantage
One reason Mantic performed strongly was its ability to avoid simply following the consensus.
Humans are influenced by other humans.
When most people believe something is likely or unlikely, individuals may become reluctant to take the opposite position.
This can create herd behaviour.
Mantic co-founder and CEO Toby Shevlane said the AI system showed less tendency to follow that pattern during the competition.
That can be valuable because a popular forecast is not necessarily an accurate one.
The Shakira Prediction Became an Example
One competition question involved whether Shakira's song “Dai Dai” would overtake “Waka Waka” on the Billboard Hot 100.
Human participants overwhelmingly expected that it would not happen.
Mantic did not place as much confidence in that consensus.
The consensus ultimately proved wrong, according to Shevlane's account of the competition.
It is a relatively lighthearted example, but it illustrates a serious forecasting principle.
When everyone reaches the same conclusion, an independent model may have value if it can identify evidence the crowd is underweighting.
The System Also Forecast Political Outcomes
Mantic's forecasting extended into politics.
Early in the tournament, the system assigned Abelardo De La Espriella roughly a 40% probability of winning Colombia's presidential election, compared with a consensus forecast of around 30%.
De La Espriella ultimately won.
A single successful forecast does not prove that an AI system can consistently predict elections or other political events. Forecasts are probabilistic rather than guarantees.
But repeated performance across many questions is precisely what forecasting competitions are designed to measure.
The Startup Was Founded Only Two Years Ago
Mantic was co-founded in 2024 by Toby Shevlane and Ben Day.
Shevlane previously worked as a research scientist at Google DeepMind. He told Reuters that the idea grew partly from his need to anticipate global developments relevant to artificial intelligence.
That background highlights an increasingly common startup pattern.
Researchers from major AI laboratories are leaving established companies to build specialised businesses around particular applications of advanced models.
Instead of competing directly with companies spending billions of dollars training frontier models, these startups can focus on applying those models to specific commercial problems.
$25 Million Is a Large Seed Round
A $25 million seed investment is substantial.
Seed rounds traditionally help young startups build an initial product, hire a small team and find their first customers.
AI has changed the economics.
Companies working with sophisticated models can require significant spending on engineers, computing infrastructure and data.
Investors are therefore committing larger amounts of capital much earlier when they believe a startup has potentially valuable technology.
Mantic's performance in a public forecasting competition provided investors with something especially useful: measurable evidence.
Microsoft Is Among the Backers
Microsoft's venture arm M12 participated in the financing alongside Radical Ventures, Thinking Machines Lab, Balderton Capital and others.
For a young startup, backing from established technology investors can provide more than capital.
Investors can help recruit employees, introduce enterprise customers and provide access to broader technology ecosystems.
Radical Ventures partner Aaron Rosenberg has also joined Mantic's board.
Businesses Are Already Interested
Mantic says its technology has attracted interest from companies and government agencies around the world.
Some organisations have already integrated the company's forecasting AI, although Mantic has not publicly identified those customers.
The potential applications are extensive.
A manufacturer might forecast supply-chain disruption.
A retailer might estimate demand.
An energy company could assess commodity risks.
A pharmaceutical company might evaluate the probability of regulatory or scientific developments.
Governments could use forecasting tools for economic planning or risk analysis.
The commercial opportunity depends on whether the system can remain accurate outside controlled competitions.
Hedge Funds See an Obvious Opportunity
One industry has an especially clear incentive to improve forecasting: finance.
Rosenberg told Reuters that hedge funds and trading firms have shown particular interest in Mantic's technology.
The reason is straightforward.
Financial markets constantly price expectations about the future.
Investors forecast inflation, company earnings, elections, interest rates, commodity prices and consumer behaviour.
Even a small improvement in forecasting accuracy could potentially have significant financial value when applied across large portfolios.
That makes investment firms a natural early customer base for predictive AI.
Forecasting Is Different From Generative AI
Most consumers currently associate AI with generation.
A chatbot writes text.
An image model produces pictures.
A coding agent generates software.
Forecasting systems operate differently.
Their value depends on calibration.
If a system says an event has a 70% probability of occurring across many predictions, events assigned that probability should happen approximately 70% of the time.
That means a useful forecasting system needs more than confident answers.
It needs probabilities that correspond meaningfully with reality.
Being Certain Can Actually Be a Weakness
A model that confidently predicts every event as either 0% or 100% can look impressive when correct.
But it performs badly when wrong.
Good forecasting involves recognising uncertainty.
A 60% forecast means an event is considered more likely than not, but there is still a substantial chance it will not occur.
That way of thinking is useful for businesses because executives rarely operate with complete information.
They must make decisions despite uncertainty.
AI Could Become a Second Opinion for Executives
One of the most interesting uses of forecasting AI may not be replacing human decision-makers at all.
It may instead provide an independent second opinion.
Imagine executives considering whether to expand into a new market.
Human analysts prepare their forecast.
An AI system independently analyses available information.
If both reach similar conclusions, confidence may increase.
If their forecasts differ sharply, executives know where further investigation may be needed.
That could make forecasting AI particularly valuable as a challenge mechanism against groupthink.
But Predictions Can Still Fail
Mantic's tournament success should not be interpreted as evidence that AI can see the future.
Forecasts depend on available information.
Unexpected events happen.
Historical relationships can break.
Political leaders can change decisions.
Economic shocks can appear suddenly.
Models can also inherit biases from the data and assumptions underlying them.
A system that performed extremely well in one forecasting environment may perform differently when deployed against real-world commercial decisions.
The true test will therefore be sustained performance over time.
Government Use Would Raise Important Questions
Interest from government agencies could create additional opportunities but also more complicated questions.
Forecasting systems might eventually influence decisions involving economic policy, national security or public planning.
That makes transparency important.
Decision-makers need to understand what information a system considered and how uncertain its prediction remains.
AI should not turn probabilistic analysis into artificial certainty.
A 70% probability is still not a guarantee.
The Startup Opportunity Is Moving Up the AI Stack
Mantic also demonstrates where entrepreneurs may find opportunities as the AI industry matures.
Building a frontier model requires enormous amounts of capital and computing power.
Only a relatively small number of organisations can compete at that level.
But startups can build businesses on top of those models.
They can specialise them for medicine, finance, law, science, forecasting or industrial applications.
This creates an ecosystem in which the biggest AI laboratories provide foundational technology while smaller companies focus on solving specific problems.
Performance May Matter More Than Model Ownership
That creates an interesting competitive question.
Does a startup need to own the underlying AI model to build a valuable AI company?
Mantic suggests the answer may be no.
If a company can take existing frontier models and consistently make them better at a commercially valuable task, that specialisation itself can become a product.
Customers may care less about who trained the original model than whether the final system delivers better decisions.
Entrepreneurship Is Becoming More Experimental
The Mantic story also shows how startup validation is changing.
Instead of simply telling investors that its technology was superior, the company competed in an external forecasting tournament.
Its performance could be compared with humans and other AI systems.
That gave investors a measurable signal before the funding round.
For founders building technically complex businesses, credible external benchmarks can be extremely powerful.
They transform a marketing claim into something investors can evaluate.
The Bigger Question: Can AI Become Better at Judgement?
Generative AI has already demonstrated that machines can produce useful text, images and software.
Forecasting raises a different question.
Can machines become better at judgement under uncertainty?
That capability could ultimately be more economically significant than generating content.
Businesses constantly make decisions based on incomplete information.
Which market will grow?
Will demand increase?
Will a competitor launch a product?
Will regulation change?
Will a project finish on schedule?
Companies spend enormous amounts of money trying to answer questions like these.
What Happens Next?
Mantic now has $25 million in new capital to develop its technology, expand its team and pursue customers interested in probabilistic forecasting.
The company's tournament performance has given it early credibility.
The next challenge is considerably harder: proving that forecasting success can translate into dependable commercial value.
If specialised AI systems can consistently improve how companies evaluate uncertain events, forecasting could emerge as another major category of enterprise AI.
For entrepreneurs, the lesson is significant.
The next valuable AI startup may not need to build the world's biggest model.
It may simply need to make existing intelligence exceptionally good at one problem businesses are willing to pay to solve.