"$25 → $4,237 in one night" with a Claude agent: no wallet, and a proof video that contradicts itself
A viral X post says a Claude-built Polymarket agent turned $25 into $4,237 overnight, and offers it free to anyone who comments, reposts and follows. It names no wallet, and the dashboard in its video breaks its own arithmetic.
The claimClaude analyzed 10,000 Polymarket wallets and built an autonomous agent that turned $25 into $4,237 in one night across 94 trades, sized with Kelly; comment 'AutoPilot', like, repost and follow to get it free. — X (@david_marco45)
The short version
The post, published October 24, 2026, tells a short story: the author lost a job, asked Claude to "analyze every top Polymarket wallet from the last 90 days and build me something", deployed the result at 11:47 PM and woke up to $25 → $4,237 from 94 trades. It describes an agent that scanned 10,000 wallets, found 7 traders with real edge, reads live news, "exploits arbitrage across outcomes" and sizes every position with Kelly. It ends with an offer: comment "AutoPilot", like and repost, and follow the account to receive it by direct message, free for 24 hours. When we fetched it, the post had about 69,600 views, 662 likes and 345 reposts.
There is nothing in the post we can check against Polymarket's records. It gives no wallet address, profile or trade link. Every Polymarket trade is public, so a real $4,237 night would be easy to show; the only evidence offered is a 24-second screen recording of a dashboard titled "AI Agent Orchestrator".
That recording does not hold together. Its clock reads 11:47 with 8 hours 24 minutes of uptime, so the session began around 3:23, not at 11:47 PM as the post says; a run started at 11:47 PM would show about 8:11 after that much uptime. During the 22 seconds shown, total profit rises from $4,237.18 to $4,388.41, but the return stays at +16,848% (correct only for the first figure) and the trade count stays at 94. It lists a maximum drawdown of $48.20 next to individual fills that lost $75.01 and $70.04. One of its eight agents shows roughly 200 trades on its own, against 94 in total. Its execution log books profit and loss on lines such as "screenshot(polymarket) → captured", and many tool calls are Solana and Binance actions (Phantom swaps, Helius wallet monitoring, Binance funding rates), not Polymarket trades.
The headline numbers also fail on their own terms. Growing $25 to $4,237 in 94 trades at the dashboard's 68.4% win rate means about 64 wins and 30 losses. At 50¢ prices that needs about 21% of the bankroll on every trade. At that size, any loss taken after the balance passed about $230 would be bigger than the stated $48.20 maximum drawdown. So the figures only fit if almost every loss came in the first few minutes, and the chart shows a steady climb from the start. The 68.4% win rate is itself a recycled number: in May 2026 LetsDataScience, citing CryptoBriefing, reported community Claude-bot claims clustered at 68.4% and a circulated promotional PDF for a 68.4% bot that contained fabricated screenshots.
Would the strategy described work? Each ingredient is something we have tested live on Polymarket at prices we could actually trade. An LLM reading breaking news to find mispricing went 3 wins and 15 losses, −70%. Arbitrage across a market's outcomes mostly disappeared at the live order book, and what remained was smaller than fees. Copying wallets selected for past results did not hold up out of sample: the top-ranked cohort returned −9.9% going forward. Verdict: an engagement post built around a product giveaway, whose only evidence is a dashboard that contradicts itself. We cannot prove no such wallet exists, because none is named; the burden is on the claim, and it offers nothing checkable.
Technical appendix
Exhibit A What the video shows, frame by frame
We downloaded the 24-second video attached to the post (720×720) and read frames at 1 s and 23 s at full resolution. Small text is blurry in places; the figures below are the ones that are clearly legible.
| Field | At 1 s | At 23 s | Problem |
|---|---|---|---|
| Clock | 11:47:21 | 11:47:43 | with the uptime below, the session began about 3:23, not 11:47 PM |
| Uptime | 08:24:14 | 08:24:36 | — |
| Total P/L | $4,237.18 | $4,388.41 | +$151 in 22 seconds, with no new trades counted |
| Return | +16,848% | +16,848% | should read +17,454% at $4,388.41 on $25 |
| Trades (header) | 94 | 94 | one agent alone lists roughly 200 trades |
| Win rate | 68.4% | 68.4% | a figure already circulating in Claude-bot promotions |
| Max drawdown | −$48.20 | −$48.20 | the fills panel shows single losses of $75.01 and $70.04 |
The "top wallets copied" panel shows five addresses, each cut off after about a dozen characters, so none can be looked up.
Exhibit B The execution log is not a Polymarket trading log
Lines legible in the log at 1 s, with the profit or loss the dashboard attaches to each:
| Agent | Tool call | Result |
|---|---|---|
| pump-sniper | Polymarket · screenshot(polymarket) → captured | +$6.30 |
| news-edge | Polymarket · screenshot(polymarket) → captured | −$5.40 |
| spread-farmer | Polymarket · screenshot(polymarket) → captured | +$10.40 |
| arb-scanner | Phantom · signTx(swap 2.1 SOL) → confirmed | +$31.20 |
| copy-trader | Binance · getFundingRate() → 0.018% crowded | +$31.20 |
| pump-sniper | Helius · monitorWallet(0xab3) → +$20k detected | +$22.10 |
Taking a screenshot or reading a funding rate does not make or lose money, and Phantom, Helius and Binance are Solana and centralized-exchange tools. Agent names such as pump-sniper, liq-hunter and dex-rebal describe crypto-token trading, not prediction markets.
Exhibit C Why the numbers cannot all be true
Taking the dashboard's own figures at face value: $25 seed, $4,237.18 profit, 94 trades, 68.4% win rate, so about 64 wins and 30 losses, sized with a fixed fraction of the bankroll as Kelly sizing does.
| Typical entry price | Fraction of bankroll needed per trade | Balance above which one loss exceeds $48.20 |
|---|---|---|
| 40¢ | 8.7% | $552 |
| 50¢ | 21% | $229 |
| 60¢ | not reachable at this win rate | — |
The final balance does not depend on the order of wins and losses, but drawdown does. A maximum drawdown of $48.20 is possible only if nearly all 30 losses happened before the balance reached a few hundred dollars. The chart in the video rises steadily from the start, and the fills panel shows losses larger than $48.20. The arithmetic is ours, from the figures on screen.
Exhibit D Our own tests of the same ingredients
Each part of the described agent maps to something this site's system has run live at executable prices:
- Reading the news with an LLM to spot mispricing: 3 wins, 15 losses, −70% at the prices we could have filled (Lab note).
- Arbitrage across outcomes: most flagged gaps vanished at the live order book; survivors were mostly stale quotes, and the rare real ones were smaller than fees (Lab note).
- Finding traders whose edge "wasn't luck" and copying them: 44 wallets with more than +15% lifetime return ran −9.9% going forward (Lab note); copying whale bets in real time netted +2.1% on $9,413, with one good month and one bad (Lab note).
None of that proves the poster has no profitable bot. It shows that the recipe described is not, on its own, a known route to profit, let alone 169× overnight.
Exhibit E How the offer works, and a disclosure
The post asks readers to comment a keyword, like, repost and follow so the author can send the bot by direct message, "free for 24 hours". Each of those steps raises the post's reach; the deadline adds urgency. We did not request the bot, so we cannot say what is sent. Anything delivered that way asks for access to a trading wallet; treat it with the care you would give any unsolicited software that can move your money.
Disclosure: this check, like the rest of the site, was drafted with help from Claude, the AI product the post names. The post's claims were checked against its own video and against our trading records, not against anything Claude said.
Source for the 68.4% figure: LetsDataScience, May 17, 2026, citing CryptoBriefing.