A Hobbyist Bot Wins the Metaculus Forecasting Cup

A bot named laertes, built by Texas hobbyist Jeffrey Liang, won the Summer 2026 Metaculus Cup on September 5, the first AI to take the seasonal title. Other bots took second and fifth. Humans took third and fourth.

The people in the field still split a $5,000 cash prize. The larger returns sat on Kalshi, a regulated prediction market, where firms that did not take the Cup have been betting real money against prices their models think are wrong.

Laertes Wins the Summer Metaculus Cup

Metaculus confirmed the result on September 17, naming Liang as the builder of laertes and congratulating Ben Shindel as the top human finisher. Liang, who has called himself a polymath, said he spent less than 150 hours and a couple of thousand dollars on computer time and data. That budget beat four startups that have raised more than $15 million between them.

For the first time, an AI forecaster has won the seasonal Metaculus Cup.

The Economist has covered the story; link in replies. Congratulations to Jeffrey Liang for building laertes, the winner in question. pic.twitter.com/akGu0IJLwT

— Metaculus (@metaculus) September 17, 2026

Yann Riviere, a forecaster at London startup Mantic, said laertes and Mantic finished first and second. The pattern matches the official ranking: machines in first, second, and fifth, people in third and fourth. Mantic, founded by Toby Shevlane, a former Google DeepMind researcher, and Ben Day, raised $4 million in pre-seed from Episode 1, with DRW and angels in the round.

The Cup is a four-month contest on Metaculus, a public forecasting platform. Entrants post probabilities on current-events questions and get scored on how close they stay to the final answer, and for how long. Bots can refresh those numbers around the clock. People cannot. Metaculus itself has noted that 24/7 automatic updates make it easier for bots to place in this format. The mechanical edge is coverage, not a mystic read of next year.

Fifty-Eight Questions Over a Summer

The $5,000 seasonal Cup prize pool sat on 58 questions. The season closed on September 2, the last items resolved on September 4, and winners went up on September 5. Guest human Dylan Matthews, a journalist, ran a side match against the Community Prediction; the community beat him. An extra $2,500 had been on the table for that result. The page still lists the pool at $5,000, and that is the cash humans shared.

One of the questions was the contest itself. “Will a bot beat all humans in the Summer 2026 Metaculus Cup?” resolved Yes, with 129 people willing to put a number on their own obsolescence.

HOW A FEW CUP QUESTIONS LANDED

• Bot vs humans: A bot did beat every human in the Summer Cup.
• Data-center curb: At least one U.S. state or EU member restricted large-scale data-center grid hookups before September 2026.
• Guest match: The Community Prediction beat Dylan Matthews.
• Copper: August futures peaked at $6.75 a pound; the community had $6.79.
• Cyclosporiasis: U.S. hospitalizations before September came in at 1,056; the community had 1,060.
• Ukraine: The Russia-Ukraine war from July 20 to September 1 stayed in stalemate, not a breakthrough.

Those items are a mix of policy, prices, disease counts, and a war. That is the point of the Cup. It is not a weather model trained on one data feed. The bots read news, filings, and markets the way a human desk would, then write down a probability and a reason, only faster and on every question at once.

Kalshi Paid Out What $5,000 Could Not

Liang’s couple of thousand dollars bought a title. It did not buy the richest year in this trade. Preseen, a forecasting startup whose bots did not make the Summer top five, publishes a Kalshi run that started at $35 and reached $1.94 million, the sixth-best return on that exchange’s all-time board. Aditya Agarwal of South Park Commons, which led Preseen’s seed round, said Theo Summer ran that Kalshi book over seven months before the firm existed in its present form. Venia Veselovsky, Summer’s co-founder, has said the same on the record.

FutureSearch, founded in August 2023 by Dan Schwarz and Lawrence Phillips after stints at Metaculus and Google, has been trading a live Kalshi book since June 9. Its public ledger shows a real Kalshi account at $105,910 on August 11, from a $100,000 start, a gain the firm labels +5.9%. That is a few months, not a seven-month moonshot, and it is still larger than the Cup’s prize. A separate simulated Polymarket book at the same shop was down 19.6% over the summer, so the edge is not automatic once the venue changes.

WHERE THE CASH ACTUALLY MOVED

Preseen’s public record is broader than one trade. Its Chestnut bot was the first bot in the Spring 2026 Cup and third of 1,283 overall. It won Metaculus’s Fall 2025 bot tournament, first of 164, and in 26Q2 it won Market Pulse outright, first of 112, humans included, the first time a bot had taken a Metaculus human field. None of that put it on the Summer Cup podium. Contest scoring and market scoring are different exams.

What Superforecasters Cost a Week

The modern craft of forecasting, as opposed to racing tips, is built on numbers you can score. In 2015, Barbara Mellers of the University of Pennsylvania and colleagues wrote in Perspectives on Psychological Science that superforecasters, the best of the Good Judgment cohort, could tell events 300 days out as well as ordinary forecasters could at 60 days. Mellers later became a scientific adviser to the Forecasting Research Institute.

A human superforecaster package can cost more than $10,000 and take a week. FutureSearch sells the same shape of answer for a few dollars and about ten minutes. Writer Scott Alexander, after running a question through the product, said the job took five minutes and cost $8 in credits. That price gap is the quiet wrecking ball. Institutions that paid for a named human desk now have a meter they can run overnight.

THE OLD PRICE OF BEING EARLY

• 300 days: Superforecasters in 2015 matched ordinary skill at 60 days.
• $10,000 and a week: Typical cost and wait for a human superforecaster job.
• Ten minutes, a few dollars: FutureSearch’s public pitch for one question.
• $8 and five minutes: What Scott Alexander said one run actually billed.

The bots are not a separate species of model. They sit on the same large language models that power the rest of the boom, then add research tools, extra data, and, at the better shops, several models arguing with each other. Mantic feeds sources that sit behind paywalls. Preseen scores pundits’ track records before it lets their copy into a forecast. Owners can also wipe a bot’s memory, feed it a snapshot of old news, and replay a past question until the prompt stops losing. No human desk can do that to itself.

The Quiet Test Where Humans Held On

The Cup is not the only exam. Metaculus runs FutureEval to line bots up against its Pro Forecasters under tighter rules. On September 2 it published the Spring 2026 write-up under the title “Pros Beat Bots, but the Gap is Nearly Gone.” That is the other true sentence in this story. A bot can win a public seasonal Cup and still trail a hand-picked human team on a shared question set built to measure the gap.

On July 16 the Forecasting Research Institute said several systems on ForecastBench, its live contest against human superforecasters, were already statistically indistinguishable from superforecasters. Cassi AI led that board. Submissions from xAI and Google DeepMind sat in the same band. Cassi also ranked above the superforecaster median on market-style questions, a first. FRI was careful: last human elicitation was in 2024, confidence bands overlap, and the clean reading is parity, not a rout.

Liang’s bot is also in Metaculus’s Summer FutureEval, a bot-only field with a $50,000 prize pool and winners due on November 5. Inference credits are covered. The dollar gap between that purse and the Cup’s $5,000 tells you which audience Metaculus thinks is building the product.

Wipe the Memory and Replay Last Year

The extra tricks are real, and they may not be why laertes won. Liang beat shops that buy data, hire researchers, and raise venture rounds. If scaffolding and paywalled feeds were decisive, the funded teams should have had the title. They did not. A single builder with a tight loop, cheap tokens, and obsessive scoring still took the seasonal Cup.

laertes and Mantic finished #1 and #2

Yann Riviere, forecaster at Mantic, on X, September 17, 2026

Riviere said he is trying a stronger centaur mix, human plus bot, in the Fall Cup, hoping to move from the top 1 percent to first. That is the near-term truce. People still collect what only a person can collect, a source who will talk, a document that is not online, a feel for a room. Bots then chew more of that input than one analyst can hold. Riviere already treats Mantic’s system as another colleague on the desk. As the models improve, that colleague will do more of the writing.

FROM THE 2015 STUDY TO THE CUP

1. 2015: Mellers and colleagues show superforecasters can work 300 days out at the level ordinary forecasters hit at 60 days.
2. August 2023: FutureSearch is founded to put that skill into software.
3. Fall 2025: Preseen’s bot finishes first of 164 in a Metaculus bot tournament.
4. Spring 2026: Preseen-Chestnut finishes third of 1,283 in the Cup, first among bots.
5. July 16, 2026: FRI reports several AI systems at parity with superforecasters on ForecastBench.
6. September 5, 2026: laertes wins the Summer Cup; bots also take second and fifth.
7. September 17, 2026: Metaculus names Liang and laertes in public.

The timeline is short. Competent AI forecasters have not been around long enough to test multi-year calls with the same rigor. Riviere still gives humans the edge on long horizons. That claim is a hunch until more years resolve.

How Far the Edge Still Holds

Weather models go random about 15 days out because the air is chaotic. If forecasting bots keep getting better, they may map a similar wall for politics, war, and prices: the share of the future that is unknown because nobody looked, versus the share that cannot be known. That is a stranger prize than a seasonal cup.

The Fall 2026 Metaculus Cup is already open. It started on August 28, closes on January 1, again with $5,000, and this season the community faces Kiko Llaneras. Riviere will try to beat the bots with a tighter human-machine loop. Liang’s cheap stack will be in the field, or something like it will. The $5,000 will again be the least interesting number on the board.

Disclaimer: This article is news reporting and analysis of a forecasting contest and of public track records on prediction markets. It is for information only and is not investment advice, trading advice, or a recommendation to buy or sell any contract on Kalshi, Polymarket, or any other venue. Readers who are considering a stake in prediction markets or related funds should speak with a licensed financial adviser who can review their own situation. Figures and standings here reflect the company pages, contest records, and posts cited, and live books can move after those dates.