Twenty-five million dollars. That is what Mantic, a London-based AI startup, pulled in as seed funding on September 18, 2026, after its forecasting system finished ahead of every human competitor in the summer 2026 Metaculus Cup. The detail I keep coming back to, though, is the ranking: Mantic came second overall. The only thing that beat it was another bot.
I build bots for a living. I read that result and did not think “AI beats humans.” I thought “the top of the leaderboard is now a machine-versus-machine bracket, and the humans are in the consolation round.” That is a different story, and for anyone writing prediction code, it is the more useful one.
Why forecasting is a brutal benchmark
Most benchmarks we throw at language models are generous. Multiple choice questions have an answer key sitting in the dataset. Coding challenges have test suites. Even agent evaluations tend to run in sandboxes where the grader was written by the same people who wrote the task.
Forecasting does not work that way. You assign a probability to a future event — political, economic, cultural — and then reality shows up and grades you. There is no partial credit for sounding confident, and no way to pattern-match your way to a score because the answer literally does not exist yet when you commit to it. The 2026 Metaculus Cup questions spanned exactly those categories, according to the published results, which means Mantic’s system had to be decent across wildly different domains rather than overfit to one.
That is the part I find genuinely hard from an engineering standpoint. A bot that is good at forecasting has to be calibrated, not just smart. Saying 70% has to mean something close to seven-in-ten across hundreds of predictions. Most of the pipelines I see builders ship, including plenty of my own early ones, are terrible at this. They produce confident prose and numbers pulled out of the model’s vibes.
What this probably looks like under the hood
Mantic has not published an architecture diagram, and I am not going to invent one. But the shape of the problem constrains the solution space enough that we can reason about it as builders.
Any system doing well on open-ended forecasting questions has to solve at least four things:
- Retrieval that stays current. A model’s training cutoff is the enemy of forecasting. Anything scoring well is pulling live information and reconciling sources that disagree.
- Base rates before reasoning. The classic human forecasting discipline is to find the reference class first and adjust second. Bots that reason first and anchor never tend to swing wildly.
- Aggregation. Single-shot predictions are noisy. Ensembling multiple runs, multiple prompts, or multiple models and combining the distributions is the cheapest accuracy win available.
- Scoring discipline. If you are not tracking a proper scoring rule like Brier or log score against resolved outcomes, you are not building a forecaster. You are building a text generator that mentions percentages.
None of that requires a $25 million budget to experiment with. That is the part worth sitting with. The core loop — question in, research, base rate, ensemble, calibrated probability out — is something you can prototype in a weekend and then spend months actually tuning. The tuning is where the value lives, and it is unglamorous work: building a resolved-question dataset, plotting calibration curves, finding out your bot is systematically overconfident in the 80-90% band.
Who is buying, and what that tells you
According to the reporting, Mantic drew interest from global companies and government agencies, with hedge funds and trading firms showing particular appetite for its forecasts. That customer mix is informative. Trading firms do not buy narrative. They buy edges that survive contact with a profit-and-loss statement, and they are unusually good at measuring whether a probability estimate was any good.
For builders, the signal is that calibrated probability output is a product, not a feature. Most bot projects I see stop at the summary — “here is what the news says about X.” The thing people with money want is the number, plus a track record that makes the number trustworthy.
The practical takeaway
If you have been looking for a project that forces you to get good at retrieval, evaluation, and uncertainty all at once, build a forecasting bot. Pick a set of open questions from a public tournament, commit your probabilities before they resolve, and keep score honestly. You will learn more about agent design from one calibration plot than from a month of prompt tweaking.
And when you are ranking your results, remember where the ceiling moved. Beating the humans got a startup a $25 million seed round. It did not get them first place.
🕒 Published: