Cash Machine
Polymarket is a site where people bet on real-world questions, like "Will this bill become law this year?" Cash Machine is an AI that places small bets there with real money. It runs on its own around the clock: no human picks, approves or places any bet. These are its balances, live.
Put in $10.11. Worth $13.38 today, +$3.27. The market puts the chance it doesn't happen at 98%.
- If right
- pays $13.70 +$3.59
- If wrong
- pays $0.00 −$10.11
Put in $10.00. Worth $10.50 today, +$0.49. The market puts the chance it doesn't happen at 96%.
- If right
- pays $10.99 +$0.99
- If wrong
- pays $0.00 −$10.00
Put in $10.05. Worth $9.94 today, −$0.11. The market puts the chance it doesn't happen at 86%.
- If right
- pays $11.62 +$1.57
- If wrong
- pays $0.00 −$10.05
Put in $9.99. Worth $9.93 today, −$0.07. The market puts the chance it doesn't happen at 76%.
- If right
- pays $13.15 +$3.16
- If wrong
- pays $0.00 −$9.99
Dashed lines mark changes to the rules for picking bets. Hover or tap one to read it.
A bet is held until the question is settled. If it was right, each share pays $1.
| Question | Worth | Profit |
|---|---|---|
| Clarity Act (H.R.3633) signed into law in 2026? ongoing | $13.38 | +$3.27 |
| Kharg Island no longer under Iranian control by December 31? ongoing | $10.50 | +$0.49 |
| Will Robert F. Kennedy Jr. leave the Trump administration before 2027? ongoing | $9.94 | −$0.11 |
| Russia x Ukraine ceasefire agreement by January 31, 2027? ongoing | $9.93 | −$0.07 |
| Will the US confirm that aliens exist by September 30? finished | $4.96 | −$0.04 |
- Look. Every hour it scans a few hundred open questions on Polymarket and skips the ones not worth a closer look.
- Research. For the most promising few, it searches the web with Exa and other search tools, reads the news, and writes a case for a bet backed by at least two sources.
- Review. AI models from different companies judge the case. Every one of them has to agree, or there is no bet.
- Bet. If they all agree, it bets $10.
The agent runs on Hermes and does every step itself, with no human in the loop. The judges almost always say no. So far it has researched 1343 ideas and placed 5 bets.
| Judge | Said yes |
|---|---|
| GPT-6 Astra | 3 of 375 |
| Claude Fable 5 | 77 of 221 |
| Kimi K3 | 100 of 202 |
| Claude opus 5.5 | 56 of 135 |
| Grok 4.7 | 1 of 64 |
Imagine it had put $10 on every idea it turned down, at the price on offer that day. 284 of them can be checked against today's prices. Those $2840.00 of bets would be worth $2948.34 now: +$108.34.
- Bets it skipped
- 284
- Would be worth
- $2948.34
- Profit if it had bet
- +$108.34
- Already settled
- 110 (46 would have won, +$105.42)
Most questions have not been settled yet, so this moves with the market every day. Older ideas are left out when the record does not say exactly which market it was about.
Every idea it passed on is kept, with its sources and reasoning.