Radical Ventures recently announced our lead investment in Mantic‘s $25M seed round. Mantic is building AI forecasting systems that now outperform the best humans at predicting the future, systems that are helping financial institutions and large enterprises make their most important decisions.
Every major decision rests on a forecast. An investment committee evaluating an acquisition, a trader assessing a position, or a GM planning a new market launch is ultimately betting on what happens next. Yet organizations still rely heavily on consultants and internal experts whose views often arrive as qualitative judgments — “likely,” “possible,” “a real risk” — that are difficult to price and rarely scored against reality.
Mantic’s forecasting engine is purpose-built for this task: the system synthesizes enormous amounts of qualitative and quantitative information from around the world into well-calibrated probabilities in real time. It breaks each question into structured sub-problems, as human experts do, and then runs specialized research agents to pursue several lines of inquiry in parallel, ultimately combining their findings into an actionable output. The team uses a range of frontier models, including ones that they have fine-tuned with reinforcement learning on forecasting tasks, and backtests the system against years of historical questions, as quantitative analysts do. Their partner, Thinking Machines, has featured this work as best-in-class on its research blog.
Customers use Mantic to track the outcomes that matter most to their business. The system continuously scans for new developments that could affect those outcomes and updates probabilities in light of new evidence. Each forecast links to a full, dynamic research report that users can access through Mantic’s web app, email, API, or MCP. Some of the world’s leading hedge funds already use Mantic to anticipate market-moving events in geopolitics, macroeconomics, and business. Fortune 500 companies are adopting Mantic to inform M&A strategy and product roadmaps.
Mantic demonstrated its state-of-the-art predictive capabilities this summer, outscoring all 676 human forecasters in the Metaculus Cup, a preeminent forecasting competition. Participants compete by assigning probabilities to roughly 60 real-world events, with accuracy rewarded over time. As recently as last year, experts suggested that AI surpassing the best human forecasters could still be a decade away.
Toby Shevlane and Ben Day founded Mantic after meeting at the University of Cambridge more than seven years ago. Most recently, Toby was a Senior Research Scientist at Google DeepMind, where he co-led work evaluating frontier model capabilities. Ben previously led research at Foresight Data Machines, where he applied AI to control systems in heavy industry. The founders have built a world-class London-based team with backgrounds from DeepMind, Palantir, Citadel, and Goldman Sachs.
We are thrilled to lead this round alongside Balderton, Thinking Machines, and DRW, among others. Aaron Rosenberg joins Mantic’s board of directors and Rich Kotite joins as a board observer. We are excited to invest in and partner with Toby, Ben, and the entire Mantic team.
Read more in Mantic’s announcement of the round. Mantic is hiring across AI research, engineering, product, sales, and operations, and is onboarding new customers, especially traders in financial markets. Reach out at contact@mantic.com.
2026 AI Founders Masterclass
Applications for the 2026 AI Founders Masterclass and Compute Cohort close soon. Our annual four-week virtual program for AI researchers and technical entrepreneurs kicks off October 7 and runs through October 29. This year’s speakers include luminaries Yann LeCun (AMI Labs), Aidan Gomez (Cohere), and Anna Goldie (Ricursive Intelligence), among others. The program is built for researchers, students, and technical founders ready to launch a startup, with the chance to hear directly from those who have done it before.
Participants can also apply to our selective Compute Cohort, which gives accepted founders up to $350,000 in compute credits, complementary legal support, mentorship from our team, and more. Spots are limited, and the window is closing fast. Masterclass applications are due October 5, and Compute Cohort applications are due October 9. Apply now at https://radical.vc/masterclass/
AI News This Week
- Google DeepMind Exodus Sparks VC Frenzy for AI’s Next Big Thing (Bloomberg) — Researchers leaving Google DeepMind are founding a new generation of AI labs, many pursuing approaches beyond large language models, and venture investors are competing to back them. In their report the Rise of NeoLabs, Radical Partners Rich Kotite and Aaron Rosenberg, who was DeepMind’s Head of Strategy and Operations, found that Google and DeepMind have produced over 40 NeoLab founders, more than OpenAI, Anthropic, and Meta combined. Radical has backed several companies founded by former DeepMind researchers, including Discovery Loop, Ricursive Intelligence, Inherent, Latent Labs, Mantic, Periodic Labs, and Orbital Industries. Colin Murdoch, the former Chief Business Officer of DeepMind and President of Isomorphic Labs, also recently joined Radical as a Senior Advisor.
- A Rocket Supply Crunch Is Making It Harder to Hitch a Ride to Space (WSJ) — Demand for satellite launches is set to outpace rocket supply through the end of the decade, according to government and industry officials. Governments and companies are racing to deploy satellites for defence programs and for commercial uses like orbital data centers, but many new rockets are behind schedule. Launch prices are rising, and some satellite operators are responding by building their own rockets or exploring acquisitions of launch companies. SpaceX’s wind-down of Falcon 9, the world’s busiest rocket, has further tightened supply.
- The AI Agent Revolution has Moved a Big Step Closer (FT) — Personal AI agents, software that carries out tasks on a user’s behalf, are emerging as the next major consumer AI category nearly four years after ChatGPT. New entrants such as Meta’s Muse, the Grok Bot, and the startup Instinct are gaining attention, and Muse has climbed to No. 1 on US app store charts. An agent that becomes the default starting point for shopping or travel would own the customer relationship and could take a cut of each referral, which puts intermediaries like online travel agents at risk. Retailers are already split, with Amazon blocking Muse from making purchases and Walmart partnering to give it free access.
- Cheap, Powerful Models are the New AI Frontier (Axios) — AI labs are shifting their competition from raw capability to cost, releasing models that pair high-end performance with sharply lower prices. OpenAI’s GPT-6 Sol and Luna cut costs for top business customers by 50%, and Anthropic says Opus 5.5 costs about 40% less to run than Opus 5. Radical Ventures portfolio companies have long competed on efficiency, with Cohere‘s open-weights Command A+ model running on as few as two GPUs, Writer training its Palmyra X5 model for about $1 million, and Reka‘s compact Edge model processing images with roughly a third of the tokens comparable models need. Chinese open-weight models are speeding up the trend, led by DeepSeek’s V4.1 Flash, which now tops OpenRouter’s usage leaderboard. Falling per-token costs are driving more usage and higher overall AI spending.
- Research: Distilling Smaller, Stronger Byte Models (Meta FAIR/University of Washington) — Distillation, where a small model learns from a larger model’s predictions, is a common way to build capable models that are cheaper to run. Researchers distilled Llama 3-8B into 1-billion-parameter models that process text either as tokens (word fragments) or as raw bytes (a collection of 0s and 1s). To do this, they used a new method that converts token predictions into byte predictions exactly. Token models performed better at low compute budgets but plateaued, while byte models started weaker and kept improving. Scaling laws project that the distilled byte model will eventually outperform its token counterpart by up to 4% and match it with one-sixth of the training data.