Polymarket trading bots

The research was done by AI agents that open web pages. Some sites refuse them; where that happened we say so. The small numbers point to the source list at the end. Where a figure is our own sum or guess, it is marked “our estimate” and the basis is given.

The claim

These are examples of the claim, copied from our collection of 59 tweets about this way of making money¹. View counts are as collected on 6 October 2026. Dates are worked out from each tweet’s ID number. Quotes ending in ”…” continue in the original. The page judges the method, not the people who tweeted, and does not say that any of them acted dishonestly.

“>Close your laptop on Friday >Leave $2,000 running on a Claude-built Polymarket bot > Open it Monday morning > $2,000 → $9,300 > The bot was scalping 5-minute BTC Up/Down markets the entire weekend. > 15 minutes to set up. > Zero babysitting. > Pure edge. > Welcome to 2026.”

@AleiahLock, 10 April 2026, 1,367,613 views¹

“i vibe coded a bot that makes me $10k/month i don’t know how to code built it in one evening with claude …”

@Argona0x, 1 February 2026, 251,029 views¹

“ANTHROPIC’S CLAUDE UPDATE JUST HELPED A STUDENT TURN $1,400 INTO $238,000 IN 11 DAYS. He used it to build a simple Polymarket bot that scanned for mispriced markets, entered when the gap was big enough, and let the repricing do the work.”

@RoundtableSpace, 14 March 2026, 205,879 views¹

“A GUY BUILT A POLYMARKET BOT WITH CLAUDE IN ONE EVENING AND WOKE UP TO $2,314 IN PROFIT. It doesn’t predict anything. It just hunts pricing inefficiencies faster than any human can react and takes the spread before it closes.”

@RoundtableSpace, 21 May 2026, 55,099 views¹

“Scalping” means making many quick, small trades. “Vibe coded” means written by asking an AI.

The 59 tweets have 6.43 million views between them (our sum)¹. The claims are of two kinds.

• Tweets that name a Polymarket account and a profit. These can be checked against Polymarket’s public data, and the ones we checked hold up as far as the profit goes (see “The accounts the tweets name”).
• Stories about an unnamed person: “a student”, “a guy”, “my friend”¹. In the text we collected, most give no account, so we could not check them. The student story appears three times with different numbers, from $238,000 in 11 days to $730,000 in 30 days¹.

What the scheme is

Polymarket is a prediction market: a website where people bet on questions with a yes-or-no answer, such as “Will Bitcoin be higher in five minutes?” Each side is a share. A share pays $1 if it turns out right and nothing if it turns out wrong. Its price moves between 1 cent and 99 cents, so a price of 60 cents means the market puts the chance at about 60%.

A bot is a program that places these bets by itself. It connects through Polymarket’s API, the official door for programs. The tweets say an LLM (a large language model, the kind of AI behind Claude or ChatGPT) can write the program for you, even if you cannot code.

The tweets describe four ways a bot is meant to earn:

• Arbitrage. Buying something for less than it is certain to pay. If the YES share and the NO share together cost 97 cents, buying both pays $1 whatever happens.
• Speed. Seeing Bitcoin’s price move on a big exchange and betting on Polymarket before prices there catch up.
• Market making. Offering to buy a little below the going price and sell a little above it, all day, and keeping the difference, called the spread.
• Copy trading. Following wallets that have won and placing the same bets. A wallet is the account that holds a trader’s money. Polymarket runs on a public ledger, so anyone can see every wallet’s trades. Data read from that ledger is called on-chain data.

Two more words matter below. A taker accepts a price someone else is already offering, so the trade happens at once. A maker posts an offer and waits. Takers pay a fee and makers do not².

The mechanism is real, and bots are allowed. The question is whether the bot in the tweets, built in an evening by a newcomer, is one of the winners.

The arithmetic

Step 1: the growth the claims need

We turned each claim into the daily growth it needs, with each day’s gain added to the next day’s stake. These are our estimates from the numbers in the tweets¹.

In the text we collected, none of these tweets gives the number of trades, the fees paid, or the worst losing run¹. Our copies are cut off at the link and do not include images or replies.

The student story’s account. One of the student tweets carries a short link to a Polymarket profile¹. Polymarket’s data shows that wallet joined on 27 February 2026⁵⁰. Its profile profit reads $238,006 on 10 March and $726,682 on 5 April, and has not moved since 7 April⁵⁰. So the profit in the tweets matches a real account. After fees it kept about $702,000, on about $23.6 million of bets⁵⁰. The data does not show a $1,400 start, a student, or code written by an LLM. On our reading of the data, the account bet about $23.6 million in those five weeks. We cannot see deposits, so the data neither confirms nor rules out the $1,400 start⁵⁰.

Step 2: the fee on every taker trade

Polymarket’s fee page gives this formula for takers: “fee = C × feeRate × p × (1 - p)”, where C is the number of shares and p is the price². The rate is 0.07 for crypto markets, 0.04 or 0.05 for most others, and 0 for geopolitics². The page says: “Makers are never charged fees. Only takers pay fees.”²

For a crypto market, where the 5-minute Bitcoin bets in the tweets sit:

• 100 shares at 50 cents is a $50 stake. The fee is 100 × 0.07 × 0.50 × 0.50 = $1.75². That is 3.5% of the stake (our sum: 1.75 ÷ 50).
• 100 shares at 95 cents is a $95 stake. The fee is about 33 cents, about 0.35% of the stake (our estimate)². The fee is small, but so is the gain: at most 5 cents a share. One wrong bet loses 95 cents, which wipes out about 19 wins (our sum: 95 ÷ 5).

These fees arrived in 2026. Polymarket’s dated changelog records: on 5 January, “Taker Fees: Enabled on 15-minute crypto markets. Fees vary by price and peak at 1.56% at 50% probability”; on 12 February, 5-minute crypto markets “Launched with taker fees enabled”; on 6 March, fees on all crypto markets; and on 30 March, fees on nearly every category “with updated rates per category”⁴⁹.

The “1.56%” is 1.56 cents for each share. On a 50-cent share that is 3.1% of the money staked, as Cointelegraph reported at the time²⁹. Today’s peak of 1.75 cents a share is about 12% higher (our sum: 1.75 ÷ 1.5625). The changelog does not give the date the crypto rate reached 0.07.

In the text we collected, none of the 30 most-viewed tweets mentions a fee, or says whether the reader’s country may trade¹. Some were posted before 30 March 2026 and still circulate¹⁴⁹.

Step 3: the “YES plus NO under $1” trick after fees

Several tweets describe buying both sides when the two prices add up to less than $1¹. For a taker in a crypto market with each side near 50 cents, the fee is 1.75 cents a share on each side, 3.5 cents for the pair². Our estimate: the two prices must add up to less than about 96.5 cents just to break even. A gap of 1 to 3 cents now loses money for a taker, before counting the risk that only one of the two orders goes through.

Step 4: how much trading an income needs

In the text we collected, no tweet gives a profit per trade. We found two kinds of figure.

One operator’s per-trade figure. A bot operator writing under a pseudonym reports an average profit per trade (“ev”, expected value) of about 3 cents for his maker bot: “Week of Feb 10: $0.031 ev/trade (baseline)”, and the same in early March²⁸. He says it runs “at 78 trades per day”²⁸. Our estimate: 78 × $0.031 is about $2.40 a day, about $73 a month. His post shows no wallet or total, and nobody has audited the figure. It links to his own bot code, which is free; we saw nothing for sale²⁸.

What named accounts kept. Polymarket’s per-wallet data gives profit before fees, fees paid, and rebates (fee money handed back)⁵⁰. For four accounts that tweets describe as bots, this is what each kept as a share of the dollars it bet. The rates are our estimates, and the meaning of the data fields is our reading, since Polymarket does not document them⁵⁰. The accounts were picked because tweets name them; they are examples, not a range for all winners. All sums in the table are ours.

At these rates, $2,000 of starting money has to be bet about 200 to 950 times over in a month to produce $10,000 (our sums).

Fees took much of it, but not all. Before fees, UpTheBlues shows about $2.28 million of trading profit. It paid about $2.12 million in fees, 93% of that, and got about $545,000 back in rebates⁵⁰. norm1e69 shows about $247,000 before fees and paid about $187,000, or 76%⁵⁰. Neither shows any maker rebate, so both appear to trade as takers, and both still came out ahead⁵⁰. Their fees came to about 1.6% and 2.1% of the dollars they bet, below the 3.5% peak, which fits trading away from 50-cent prices (our estimates)⁵⁰. So the rule is not “only makers win”. It is that a taker’s gain per trade before fees must be clearly bigger than the fee at the prices it trades.

A wider look. On Polymarket’s crypto leaderboard for the last month, among the top 3,000 accounts by profit, the 177 that traded more than $1 million had a middle (median) profit of 0.5% of volume (the total of all bets placed); across all those showing a volume it was 3.9% (our estimates)¹⁵. Treat both with caution. That board appears to show profit before fees, its volume figure is zero for 181 of the 3,000, and a one-month list includes lucky one-off bets¹⁵. Telonex’s one-week study found the same mix: its highest profit-per-dollar wallets were either “snipers” with a handful of big bets or fully maker bots active in over 2,000 markets¹⁹.

Step 5: how big the pool is

Arbitrage, measured for one year. Researchers studied Polymarket’s on-chain data from 1 April 2024 to 1 April 2025 and put the total taken by the arbitrage strategies they measured at $39,587,585.02¹⁸. That is about $3.3 million a month for everyone together (our sum). The top ten accounts took $8.18 million, about 21% (our sum of the paper’s Table 1)¹⁸. The authors describe “some very big players with bot-like behaviour”¹⁸.

Three limits. The year includes the 2024 US election. Polymarket charged no fees then¹⁸. And only about $95,000 came from arbitrage across related markets, the kind that needs something like an LLM to notice that two questions are linked¹⁸. A newer study of one class of markets estimates $1.12 million of arbitrage profit in all³⁷.

Market making in short crypto markets, measured for one week. Telonex, a company that sells market data, analysed Polymarket’s 15-minute crypto markets for 2 to 8 February 2026: 46,945 wallets and $153 million traded¹⁹. Makers as a group ended the week up $728,501¹⁹. If every week were like that, makers together would earn about $3.1 million a month (our estimate: 728,501 × 4.3). But the top five wallets alone made $811,403, more than the makers’ combined net result¹⁹. A newcomer’s share has to be won from them. Telonex’s figures leave out fees: “Our PnL calculations are pre-fee”¹⁹. (PnL means profit and loss.) Taker fees were already charged that week, and today’s are higher⁴⁹.

Step 6: how many accounts get there

These figures cover all traders, including people who placed one small bet. When this page was first written we had found no source that counts bot operators separately. A second search found one, covering nine days in August 2026; it is under “Bots as a group” below⁶².

• Across 2.5 million wallets. An independent analyst, Andrey Sergeenkov, studied on-chain data from April 2024 to April 2026: “15.9% of Polymarket traders are in profit and 84.1% are in the red.”²⁰ Only 2% of wallets had ever made more than $1,000 in total, and 840 addresses (0.033%) more than $100,000²⁰²¹. Four months in a row above $5,000 happened for 0.015% of addresses²⁰. His caveats: “I only count realized PnL. If a trader bought tokens on an unresolved market and hasn’t sold them, my data shows that as a loss”, and “Every number represents addresses, not necessarily people.”²⁰
• In the short crypto markets. In the Telonex week, “just 36.8% finished the week in profit”¹⁹. The middle wallet lost about $3, and takers as a group lost, before fees¹⁹.
• On Polymarket’s crypto leaderboard for the last month. Rank 1 shows $740,084 profit, rank 50 $39,270, rank 1,000 $1,743, rank 3,000 $326 and rank 10,001 $29¹⁵. The list stops near rank 10,000, so we do not know how many people traded. The figures appear to be before fees, so they are upper limits¹⁵.
• Big traders lose too. Of the 50 crypto accounts with the highest volume last month, 10 show a loss, the largest $649,391 (read at 11:44 UTC on 7 October 2026)¹⁵.

Step 7: the costs

The tools are cheap, which is the part of the claim that holds up.

• Polymarket says: “There are no Polymarket fees to deposit or withdraw USDC”². USDC is a digital dollar.
• Polymarket accepts up to 5,000 order requests per 10 seconds from one internet address⁷. A separate limit of 40 orders a second for a new account was announced in July 2026 and is not yet enforced⁵¹. Either is ample for a beginner’s bot.
• Claude’s price list shows $1 to $10 per million input tokens (pieces of words) and $5 to $50 per million output tokens⁴⁵.
• One person who tried it reports paying $24 a month for a rented server³¹. Listed prices for a basic rented server run from $4 to $48 a month at two large providers⁷¹.
• Data is priced too. Past order-book data for testing a strategy costs $99 or $199 a month at Telonex. On-chain data costs $29, $69 or $199 a month at Bitquery. A sports odds feed is free for 500 requests, then $30 to $249 a month⁷¹.

The real costs are trading losses and taker fees.

What people who tried it report

Studies of all traders

Four counts between 2024 and 2026 put the share of losing accounts between 67% and 86%. They describe every trader, not bot operators.

• November 2024: Decrypt reported that “86% of accounts have a negative realized profit”²⁶.
• December 2025: data from the analyst DeFi Oasis put 70% of 1.7 million addresses in loss²⁰²⁷. Sergeenkov says that count missed one kind of cost²⁰.
• April 2026: Sergeenkov’s 84.1%, above²⁰.
• April 2026: a London Business School and Yale study, as reported in a news brief, classed 3.14% of 1.72 million accounts as skilled winners and found that 67% of accounts carried all of the net loss²⁵.

A Bloomberg analysis, as reported by two trade sites, found that “5% of wallets accounted for 75% of trading volume” and “made $131 million on their trades”, while everyone else together lost the same amount²⁴. “More than 100,000 wallets lost at least $1,000 — nearly double the number that gained the same amount.”²³ A news summary of the same analysis adds that about 800 bot-like accounts each cleared more than $100,000 between January 2025 and April 2026⁶⁶.

One study since the current fees. An earlier version of this page said we had found no study of results after 30 March 2026. A second search found one. The Pew Research Center followed 11,989 active Polymarket wallets from 7 May to 19 June 2026⁶⁸. The typical user roughly broke even: 58% ended within $100 of where they started, and the average net loss was under $2 on average spending of a little over $600⁶⁸. At the edges, “7% had a net profit of more than $1,000 over this six-week period, while 9% lost a similar amount.”⁶⁸ The wallets were drawn from traders on 10 busy events, so they are not a random sample, and Pew does not separate bots; 11% of the wallets placed 1,000 or more trades⁶⁸.

Bots as a group: ahead over Polymarket’s whole history, behind in one recent count

Correction, 7 October 2026. This section was headed “Bots as a group win”. That rested on one report read second-hand. A count made after the current fees found the opposite for nine days in August 2026, so the heading and text are changed. Both findings are below.

Galaxy Research, whole history. Galaxy, a crypto financial firm, separated likely bots from ordinary users. We first read its report through a news site²² and have now read the report itself; the figures match⁶⁴. The 125,429 accounts placing 50 or more orders per active day placed 80.8% of all orders and ended $246.8 million ahead²²⁶⁴. They are 4.1% of accounts and 41% of the money traded⁶⁴. Among the rest, “69.2% of retail Polymarket accounts finished below break-even”²².

Three limits. The data covers Polymarket’s whole history since 2020, much of it before fees⁶⁴. Galaxy gives no share of automated accounts that were each in profit⁶⁴. And it says the gain is concentrated: “A large number of accounts churn for trading rewards, and a much smaller number do genuine market making and arbitrage.”⁶⁴ To churn is to trade back and forth mainly to collect rewards.

Bitquery, nine days after the current fees. Bitquery, a company that sells blockchain data, classed a wallet as machine-run if it placed orders in all 24 hours of a day with a middle gap of under 60 seconds between orders⁶². From September 2025 to August 2026 that was 12,685 of 2,347,464 wallets, about 1 in 200. They placed 63.1% of trades and 28.3% of the money⁶².

For markets that opened and settled between 11 and 19 August 2026, 3,362 machine-run wallets together lost about $473,000 before fees and paid about $2.35 million in fees, ending $2,822,122 down⁶². Of them, 34.4% finished ahead. Of the 84,417 other wallets, 37.8% finished ahead⁶². Machine-run wallets were about 1 in 7 of the wallets that made over $10,000 and about 1 in 6 of those that lost over $1,000⁶². Bitquery’s summary: “Automation does not buy a better result here. It buys a bigger one.”⁶²

This is the first figure we found for the share of bot wallets in profit. Its limits: nine days, settled markets only, one firm’s method, and a test strict enough that slow bots count as people⁶².

Bitquery, the short crypto markets. A second Bitquery study covers the 5-minute and 15-minute crypto “up or down” markets, the ones in the tweets, for one week in August 2026 after fees⁶³. Of all wallets, 31% finished ahead. The more markets a wallet played, the bigger its typical loss: the middle wallet lost $0.96 among those that played one market, $39.61 among those that played 101 to 500, and $101.61 among the 1,554 that played more than 500⁶³. Of that last group, the one most like a bot, 32.2% finished ahead⁶³.

The best wallet that week played more than 6,000 markets in 8 days, paid more fees than any other, and still finished tens of thousands of dollars ahead⁶³. The worst was a maker that played over 15,000 markets, paid almost no fees, and lost more than the best one made⁶³. Fewer than 4 in 100 of February’s players (3.6%) were still betting in August⁶³. About $150 million of fees were charged on these markets from January to mid-August 2026⁶³.

What this adds up to. Over the whole history, automated accounts as a group took money from everyone else, and a small number of them took most of it⁶⁴. In the one recent window measured, running a bot did not raise the odds of finishing ahead; it raised the size of the result in both directions⁶².

The accounts the tweets name

We checked six named accounts against Polymarket’s own data on 7 October 2026⁵⁰. For five, the tweet’s figure is at or a little below the profit on the account’s profile page. That profile figure is before fees.

So the people who tweeted these figures quoted Polymarket’s profile pages accurately. What the profile figure hides is fees: UpTheBlues and norm1e69 each kept under a third of the headline, while BoneOhio’s rebates almost covered its fees (our sums)⁵⁰. Lucerys does not fit: its profile shows far more than its trading figures, and we do not know why⁵⁰.

One more public case resembles the tweets’ stories. The 0x8dxd profile starts on 5 December 2025 with $314 of profit and reads about $2.38 million today, before fees⁵⁰. We do not know who runs it or with what money.

In every case the data shows profit, not how it was made. It does not show that an account is a bot, that an LLM wrote it, or that a newcomer runs it.

Tests of the method in the tweets

• The Bitcoin speed bot. A researcher recorded Polymarket’s 15-minute Bitcoin markets alongside the Binance exchange and tried to trade the gap³⁵. Polymarket’s prices followed large Binance moves “after a median 347 ms”, a third of a second; another of his measures gave 16 milliseconds³⁵. His model did not beat the market’s own prices, and “simulated trading nets -0.116 normalized payoff units per attempted trade”, a loss³⁵. That is one model under one author’s assumptions. For scale: Polymarket holds taker orders in crypto markets for 150 milliseconds⁴⁹, and lets approved traders place machines in its own data-centre region⁸. A bot that asks an LLM before each trade waits seconds (our reasoning, not a measurement).
• LLMs as forecasters. In a live test of seven AI models choosing Polymarket bets, “only two of seven models achieve positive financial returns”, at 17.6% and 6.2% on the paper’s measure³⁶. This bears on bots that ask an LLM to predict events, not on arbitrage or market making.
• The 5-minute Bitcoin markets. Researchers wrote in June 2026: “Manipulators capture a large amount of profit, mostly from retail.”³⁸ The method was pushing Bitcoin’s price at the moment a market settles. In August 2026 Polymarket changed the settlement price from a single reading to a 60-second average⁴⁹. We found no measurement since. A second paper documents matched orders deliberately made to fail, with “at least $1.49M in profit” to those doing it³⁹.
• Maker bots can be picked off. CoinDesk reported in January 2026 that one trader moved the price of the XRP coin on a quiet weekend and took $233,000 from Polymarket market-making bots, at a cost of roughly $6,200⁷⁰. The bots’ loss figure comes from a data account on X, not from records we checked⁷⁰.
• What one operator says changed. The operator from Step 4 writes that in February 2026 “pure latency arbitrage strategies” in general “stopped working the same day” a delay on taker orders was removed, and that fees then sank thin-margin taker bots²⁸. Of his own bot, which posts maker offers at 70 cents and above: “Almost nothing changed for my bot.”²⁸ Polymarket’s changelog does not record the delay change he describes⁴⁹. When this page was first written, his was the one first-person report we had found of a bot staying positive through the fee changes, at about $73 a month by our sum. A second, with a public wallet, is under “The upside”⁵⁷.

Individual write-ups with numbers

Most individual write-ups we found with real money and a published total were a small loss or never properly started. The four below are self-reported. The exceptions are the operator in Step 4²⁸ and one with a public wallet, described under “The upside”⁵⁷.

• A momentum bot made 104 trades and lost $239.11, with reported equity of $4,269.91. The author writes: “the bot ran for several months and lost money every month.”³⁰
• An author gave an AI agent about $23. After about 46 hours and roughly 140 trades it had about $1.50³¹. He says the account behind the viral post that inspired him had a Polymarket referral link in its profile, and concludes that referrals were the point³¹.
• A strategist at the fund manager Purpose Investments used AI to build two strategies. Both looked profitable in backtests (trial runs on past data) and lost with real money. A $400 wallet fell to $134 in about two weeks³².
• A newsletter writer gave an AI agent $100. After three strategies failed in practice runs, it put only $2 into real trades³³.

The author of poly-maker, a free market-making bot with 1,505 stars (GitHub’s mark of approval), writes: “Market making on Polymarket is competitive and can lose money. This is a reference implementation and a research harness, not a guaranteed-profitable product.”⁴⁴

Correction, 7 October 2026. An earlier version of this page said we had found no independently checked success story from an individual who built a bot with an LLM, and that this might be a gap in our search. It was. A second search found one: a data engineer whose wallet is public and whose profit Polymarket’s data confirms⁵⁷⁶¹. That an AI wrote the code is his own statement. The case is under “The upside”.

Copying winning wallets

We found no study of what people earn by copying wallets. Three things bear on it.

• Past winners often do not repeat. The London Business School and Yale study, as reported, found that about 60% of “lucky winners” became losers in a second sample²⁵.
• The copier trades later and at a worse price. Polymarket warns “there is no guarantee that it will be possible to transact a desired amount of shares without impacting the price significantly”¹¹.
• The copier pays Polymarket’s taker fee and may pay the tool’s fee on top²¹⁴.

Ready-made bots and stolen wallets

A bot needs your private key, the secret that controls the wallet. The security firm StepSecurity reported in March 2026 that a legitimate project’s GitHub organization “has been hijacked” and used to publish more than 20 code repositories presented as Polymarket trading bots, which pulled in four booby-trapped packages that take wallet keys³⁴. The main repository showed 366 stars, “likely inflated”³⁴. The report gives no count of victims. Polymarket’s terms say it “cannot … reverse any transactions you initiate”⁹.

What the rules allow now

Bots are allowed. Polymarket publishes the API, and its guide for market makers assumes a program is doing the work: “Add a kill switch: Cancel all open orders when errors or position limits require the strategy to stop”⁶⁷.

Many readers of these tweets may not trade on the main site. Polymarket’s terms, in force from 11 August 2026, say trading is “NOT PERMITTED” for people in the United States, Alberta, Australia, Belgium, British Columbia, France, Germany, Hungary, Ireland, Italy, Ontario, Poland, Quebec, Russia, Singapore, Slovakia, Slovenia, Taiwan, Thailand, the Netherlands, New Zealand, the United Kingdom “OR ANY OTHER RESTRICTED TERRITORY”⁹. The block covers programs too: “Orders submitted from blocked regions will be rejected.”⁸ Polymarket’s own code project for AI agents says the bar applies “via UI & API and including agents”¹⁷.

The developer page lists 7 places blocked completely, 37 countries and provinces limited to closing existing bets, and 5 restricted on the website only⁸. Polymarket’s pages do not fully agree: the help centre says 39 countries¹⁰, and the terms name Hungary and Slovenia, which the developer page omits⁸⁹. Check your own country.

Getting around the block breaks the terms. “USE OF A VIRTUAL PRIVATE NETWORK (“VPN”) OR ANY SIMILAR TOOL TO ATTEMPT TO OR TO CIRCUMVENT THE RESTRICTIONS SET FORTH HEREIN IS STRICTLY PROHIBITED”, and wallets may be “PLACED IN CLOSE-ONLY MODE”⁹. A VPN hides where you are. Close-only means you can sell what you hold but not open new bets.

US residents have a separate, regulated site. Polymarket US is “CFTC-regulated”⁹; the CFTC is the US regulator for this kind of trading. It publishes an API for placing orders⁵². Its taker fee uses the same formula with a rate of 0.0695, “Effective exchange-wide from 10 AM ET, Thursday October 1, 2026”⁵². We did not check whether it offers 5-minute crypto markets, and none of the figures on this page come from it. In January 2022 Polymarket’s operator settled CFTC charges that it ran an unregistered facility for event-based contracts, and paid a $1.4 million penalty, which the CFTC said was reduced for its cooperation⁴⁷.

Some bot behaviour is banned. The terms prohibit spoofing (posting offers you do not mean to honour), front-running, collusion and wash trading, meaning orders whose purpose “is to avoid taking a bona fide market position exposed to market risk”⁹. A bot that trades against its own second wallet to collect rewards fits that.

The platform can change under a bot. On 28 April 2026 Polymarket replaced its trading system: “All existing order books will be cleared during the maintenance window.”¹² An order book is the list of open offers. Bots and tutorials written before that date need rewriting. In 2026 the fees changed several times, the taker delay twice and the settlement price for short crypto markets twice⁴⁹. Maker rebates are “at the sole discretion of Polymarket and may change over time”³.

Losses are yours. “YOU CAN LOSE UP TO THE ENTIRE AMOUNT OF THE CRYPTOASSETS SUPPLIED TO THE CONTRACT.”⁹ The terms cap the company’s liability at $100 and leave tax to you⁹. We did not check the tax rules of any country.

Regulators are watching these markets. A CFTC advisory of 25 February 2026 lists insider trading, wash sales and manipulation on prediction markets as things it can act on⁴⁸. The CFTC announced four cases against individuals between April and August 2026. Two are complaints about trading on polymarket.com; both were pending, with nothing decided, when we read them on 7 October 2026. One alleges that a US service member made “more than $404,000 in profits” from information the public did not have⁵³. The other is against an employee of a technology company⁵⁴. These are the CFTC’s allegations, not findings. Two ended in settled orders with penalties and three-year trading bans. Neither release names the platform; in one, the CFTC thanks the rival site Kalshi for its assistance⁵⁵⁵⁶. None was about an ordinary arbitrage or market-making bot. These cases concern only the people and companies named in the CFTC’s releases. We found no regulator action against any account quoted or named on this page, and we do not suggest any connection. Polymarket says it has given details of more than 315 wallets to law enforcement¹⁶. For copy traders the lesson is that some wallets worth copying may be winning on information a copier cannot have.

Who makes money from it

None of the earners below depends on the newcomer’s bot making a profit.

The platform. DefiLlama, a site that tracks crypto businesses, shows about $103 million in fees paid by Polymarket users over the last 30 days, of which Polymarket kept about $15.1 million⁴³. We did not check its method.

Large makers and the biggest traders. Part of the taker fees goes back to makers every day: 20% in crypto markets, 15% in sports and 25% in most other categories³. Makers can also earn rewards for keeping offers open; one stated pool was $1 million across short crypto markets for August 2026, now ended⁵. Takers get fees back only at size: 3% once their 30-day trading passes $2,000 on Polymarket’s count, 50% at $10 million⁴. So fees flow from small takers to large makers and very large traders. Polymarket’s data shows the scale: one maker wallet with about $140 million bet shows about $698,562 in rebates and rewards beside $809,704 of trading profit, and another shows about $875,993 beside $695,018⁶¹. The one small maker we could check received $43 in rebates and rewards on $140,154 bet⁶¹. Rewards under $1 are not paid⁵.

Tool sellers.

• Polymarket lets apps that route orders add their own fee of up to 1% for takers: “They’re additive — they stack on top of platform fees, never replace them.”¹⁴
• Kreo is a trading app for Polymarket and Kalshi that includes copy trading⁴¹. A 141,539-view tweet by an account that is not Kreo’s names the app and says “total cost: $0”¹. That is the tweet’s wording, not Kreo’s, and our copy of the tweet is cut off. Kreo’s own page lists its fees: “Fees apply on buys, sells, and redeems. Kreo charges an industry-low 0.3% fee on redeems/claims” (a redeem is collecting a winning bet)⁴¹. Its per-share formula is the same as Polymarket’s crypto taker formula, and the page does not say whether it is Polymarket’s fee passed on or a charge on top²⁴¹.
• PolyTrack, a wallet-tracking tool named in one tweet in the collection, appeared to charge $19 a month, $149 a year or $299 for life, with a free tier, when we read its site on 7 October 2026¹⁴². We read the prices from the site’s code, not its displayed page, so they may differ from what a visitor sees.

People who refer others.

• Polymarket pays referrers 10% of the fees of people they bring in, for 30 days or until the referred trader reaches the Platinum tier¹³.
• Kreo pays 30% of fees from direct invites, and: “When another Kreo user sets up a copy trade task following your wallet, you automatically earn 15% of the fees they generate on every copied trade.”⁴¹ On these terms, a trader can be paid for attracting copiers whether or not the copiers profit.
• Of the short links in the 59 tweets, 6 lead to Polymarket profile pages with an extra tag in the address (“?r=” or “?via=” followed by a name)¹. These may be referral tags. We did not confirm what they are or that they pay anyone.

A course seller. Moon Dev’s website sells “Lifetime All Access” for $1,500, including a “Polymarket Trading Bot Course”, with a “90 Day Money Back Guarantee”⁴⁰. Moon Dev also gives material away: a free tutorial video, and a tweet saying the code is on GitHub¹. We found no figures on what Moon Dev’s members earn.

What the tweets themselves offer. By the labels in our collection, 37 of the 59 tweets show nothing for sale, about 11 promote a tool or bot, and the rest link to tutorials, guides or a paid community¹. The labels were not checked against each account.

Regulators on bot sellers. We found no court or regulator action about Polymarket bot courses or tools. The nearest is a CFTC advisory from January 2024, about people who buy a bot or hand money to a bot operator in currency and crypto trading: “Scammers claim AI-created algorithms can generate huge returns—sometimes tens of thousands of percent—or yield 100 percent “win” rates.”⁴⁶ The advisory is not about Polymarket and names nobody on this page. We found no regulator statement about any seller named here. It does not say that building your own bot is unlawful or cannot work.

The upside

Some accounts do very well on Polymarket, and its public data lets anyone check them. The studies and the data agree on who most of them are: a small number of accounts that bet very large sums and keep a small share of each dollar, as makers or with mixed strategies^1518192250. Each case below says whether the profit was checked against that data or is the person’s own statement, and whether the person sells anything on the subject.

Two limits apply to every checked case. The data shows profit. It does not show who runs an account, what money it started with, or what wrote its code⁶¹. And “kept” means after fees and rebates, on our reading of data fields that Polymarket does not document⁵⁰⁶¹.

The best documented outcomes

The two largest machine-run accounts. Profit checked against Polymarket’s data. Operators unknown. We found no link to anyone selling a course or tool. Bitquery classes both as machine-run⁶².

• swisstony joined in July 2025. It kept about $21.0 million after paying about $5.05 million in fees, on about $1.0 billion bet⁶¹. That is about 2.1% of the dollars it bet (our sum). Its profile lost $24,000 and then $120,000 in its first two months. It then rose, to a gain of $8.1 million in July 2026, before fees⁶¹. It last traded on 26 August 2026⁶¹.
• RN1 joined in December 2024. It kept about $14.3 million, about 2.2% of the dollars it bet (our sum)⁶¹. Its profile gained $3,000, $48,000, $149,000 and $443,000 in its first four charted months, July to October 2025, then $0.8 million to $2.0 million a month from November 2025 to August 2026 and about $0.57 million in September 2026, before fees. It was up in all 16 months charted⁶¹.

So even the largest winners took months to grow, and one lost money first.

The top 50 of the all-time crypto board. Checked. Operators unknown. We took the 50 accounts with the highest all-time profit on Polymarket’s crypto leaderboard and looked up each wallet⁶⁰⁶¹.

• 49 of the 50 are ahead after fees and rebates across their whole account⁶¹. The one exception shows $494,547 of profit on the crypto board and is about $342,824 down overall⁶⁰⁶¹.
• The board shows $2,215,534 for rank 1 and $409,197 for rank 50⁶⁰. Those board figures cover crypto markets only and appear to be before fees.
• 26 of the 50 joined in the last 12 months, and 16 have made more than a million trades⁶⁰⁶¹.
• Most kept between 0.4% and 5% of the dollars they bet (our sums)⁶¹. At those rates the income comes from volume.

New accounts under the current fees. Checked. Operators unknown. Of the top 50 on last month’s crypto board, 44 wallets joined in 2026 and 35 joined since April 2026, after fees reached nearly every category⁶⁰⁶¹⁴⁹. The joining dates come from the per-wallet data, which returns none for two of the 50⁶¹. Of the 50, 48 are ahead after fees over their whole life⁶¹. Three with about 228,000 to 1.37 million trades each:

x-MoneyForWhiskas paid about $416,000 in fees on the way⁶¹. A fourth account, no. 7 on last month’s board, is named “dangerously-skip-permissions”, which is a setting in the AI coding tool Claude Code. It joined on 16 July 2026, has made about 481,000 trades, and kept about $131,574 after paying $163,444 in fees, on $7.07 million bet⁶⁰⁶¹. The name is only a hint. Nothing shows what wrote its code⁶¹.

This cuts against reading the 2026 fees as the end of the opportunity. Three cautions. A list of last month’s winners is picked for success and includes luck; several of the 50 have only 2 to 80 trades, which are one-off bets and not bots⁶⁰⁶¹. A new wallet is not a new person: an experienced operator can open one. And these sums were earned on $2.4 million to $18.2 million of bets⁶¹.

One small operator. Profit checked against Polymarket’s data. The rest is his own statement. He sells nothing that we saw. A data engineer who writes as “Kacho” ran a bot on sports and esports (video-game contest) markets from 6 January 2026 and named the wallet, b00k13⁵⁷⁷⁴.

• His account. 3,858 bets, about $95,830 traded, $4,973 ahead by April: $1,898 in January, $2,506 in February, $390 in March and $180 in April⁵⁷. He says he started with about $1,000⁵⁸ and that an AI wrote almost all the code: “I barely ever wrote a line of code. Nevertheless, don’t even consider this will be trivial for any AI model to code on its own.”⁵⁷ He says he has a bachelor’s degree in computer science and a master’s in AI⁷⁴.
• The check. Polymarket’s data for that wallet shows it joined on 6 January 2026. Its profile profit read $2,010 on 31 January, $4,461 on 28 February, $4,828 on 31 March and $4,962 on 28 April, which matches his account⁶¹. Today it reads $6,020. After $281 in fees it has kept about $5,786, on $140,154 bet⁶¹. That is about 4.1% of the dollars bet (our sum).
• Since then. The profile was flat from late April to mid-June, then added about $1,055 by today, roughly $285 a month (our estimate)⁶¹.
• What he sells. The code is free⁵⁹. His site has an email sign-up; we saw no paid product⁵⁷.

One maker from before the fees. His own statement, published by Polymarket. He gives his code away. The author of the free poly-maker bot told Polymarket’s newsletter: “I started with $10,000 in capital and was earning about $200 per day. As I scaled up and refined the system, I reached $700-800 daily at peak.”⁶⁷ He gave no wallet. He said most of it came from rewards for keeping offers open, and that after the 2024 election those rewards “decreased significantly, making the strategy less profitable”⁶⁷. It is a 2024 result. Polymarket has an interest in attracting makers (our view), so its newsletter is not a neutral place for the figure. His code page now says market making “can lose money”⁴⁴.

Press stories, partly checked.

• A widely shared article said a “Claude-powered trading bot” turned $1 into $3.3 million⁶⁹. The account it names, sovereign2013, is real. It kept about $3.56 million on $200.8 million bet, over 1.25 million trades, earned between October 2025 and April 2026 and flat since May 2026⁶¹. The reprint we read gives no evidence for the $1 start or for Claude⁶⁹.
• The 0x8dxd account, in Step 4, kept about $2.17 million⁵⁰. It last traded on 24 August 2026⁶¹. We could not check the small starting sum that press stories give it.
• The student story’s account, in Step 1, kept about $702,000⁵⁰. The start, the student and the AI are unchecked.

What the top groups earn

These cover all traders, not only bots, unless stated.

• Bot-like accounts. About 800 each cleared more than $100,000 between January 2025 and April 2026, in Bloomberg’s analysis as a news site summarised it⁶⁶.
• The top third of one percent. In Sergeenkov’s study of 2.5 million wallets, about 8,000 addresses (0.32%) made more than $10,000 in total²⁰. While they were active, 1.25% averaged more than $1,000 a month, about 6,600 (0.26%) more than $5,000 a month, and about 3,250 (0.13%) more than $10,000 a month²⁰.
• How long it lasts. Of that $5,000-a-month group, 53% earned it in a single month, and “Only 172 addresses out of 6,600 with average profit above $5,000 were active for more than a year.”²⁰
• The top hundredth of ordinary accounts. Among accounts trading at a human pace, Galaxy puts the top 1% at a gain of $3,381 or more over their history and the bottom 1% at a loss of $4,804 or more. The middle half sits between a loss of $36.64 and a gain of $0.40⁶⁴.
• Six weeks under the current fees. In Pew’s sample, 7% made more than $1,000 and 9% lost more than $1,000⁶⁸.
• Skill lasts. Yale’s write-up of the London Business School and Yale study says about 3% of accounts were classed as skilled, and that 44% of those classed as skilled on one half of their bets were classed as skilled again on the other half, against about 10% for mutual fund managers⁶⁵. Skilled traders and a smaller group of market makers “represent fewer than 3.5% of all accounts, yet capture over 30% of total gains”⁶⁵. The same study found about 60% of “lucky winners” reversed²⁵. We did not open the paper itself.

A realistic good result in the first year or two

This is our estimate. It rests on the two small operators who published figures, so treat it as thin.

• For a competent programmer who finds a quiet niche: a few thousand dollars in total, most of it in a few good months before faster bots arrive. Kacho’s wallet gained about $2,000 and $2,450 in its two best months, then a few hundred dollars a month or less⁶¹.
• For a simple maker bot without a special edge: tens of dollars a month. The operator in Step 4 works out at about $73 a month, by his own unchecked figures²⁸.
• Six figures in the first months has happened for new wallets in 2026, on $7.07 million to $18.2 million of bets⁶¹. Nothing shows those operators are newcomers.

For scale, rank 1,000 on last month’s crypto board shows $1,743 before fees, across all traders and including luck¹⁵. Four months in a row above $5,000 happened for 0.015% of addresses²⁰.

The other side of the same data

• Size cuts both ways. On our reading of Polymarket’s data, “suntori”, the second-biggest machine wallet by fees in Bitquery’s table, is about $11.1 million down after fees, on $199 million bet and $3.18 million of fees paid, and its profile shows no positive month⁶¹⁶².
• Winning runs end. Four crypto wallets that tweets or press describe as bots stopped trading on that wallet in early 2026, about when fees arrived: PurpleThunderBicycleMountain on 21 January (kept about $846,000), gabagool22 on 20 February (about $865,000), distinct-baguette on 25 March (about $718,000) and k9Q2mX4L8A7ZP3R on 6 April (about $1.48 million)⁶¹. Each earned its total in 2 to 5 months⁶¹. They may have moved to new wallets; the data cannot show that.
• The common outcome is still a small loss. The middle wallet in the short crypto markets lost about $3 in a week in February 2026¹⁹, and the middle heavy player lost $101.61 in a week in August 2026⁶³. The losing write-ups we found lost between $21 and $266³⁰³¹³².

What it takes to compete

What the winners have that most who try do not

1. An edge that does not depend on predicting. An edge is a repeatable reason your bets earn more than they cost. Kacho’s bot “never once tried to predict who would win”⁵⁷. It took a fair price from bookmakers’ odds, posted maker offers on Polymarket, and acted only when the gap was at least 7%⁵⁷. Galaxy’s report describes the profitable automated accounts as doing “genuine market making and arbitrage”⁶⁴.
2. A place with few competitors, or a subject you know. Kacho chose esports because the spreads were “unusually wide (sometimes 20-30¢)” and the markets still had volume⁵⁷. Among human-paced accounts, Galaxy found 36.8% of finance specialists and 41.2% of tech and science specialists in profit, against 30.4% of generalists and 25.1% of sports specialists⁶⁴. Galaxy warns that the tech and science group is small and may include insiders⁶⁴. Its comment: “concentration pays where a trader plausibly has an edge.”⁶⁴
3. Volume (the total of all bets placed), and the money to sustain it. The top 50 crypto accounts mostly kept 0.4% to 5% of the dollars they bet, and the named accounts in Step 4 kept roughly half a cent to two and a half cents per dollar⁵⁰⁶¹. Six-figure results came on millions of dollars of bets⁶¹. Kacho is the exception that shows the other route: about 4.1% on a small volume, in a niche, for a few months⁶¹.
4. Fresh data and speed. Kacho’s odds refreshed about every 10 minutes. The share of markets where his offers were taken fell from 37% in January to 1% in April. He puts this down to faster bots and to his own move into sports markets, and says he cannot separate the two⁵⁸. He was filled mainly when his price was out of date⁵⁸. The legs of his trades that were covered on both sides made $8,293; the ones left uncovered lost $3,185⁵⁷. Of his free code he writes: “the code was never the hard part. The edge was fresh odds and speed, and here both are on you … Run it as-is and you’ll donate to quicker bots.”⁵⁹
5. Programming and monitoring skill, even with an AI writing the code. Polymarket’s guide for market makers expects offers on both sides, live data feeds, limits on what the bot holds, and a kill switch. It notes that “Orders cannot be edited in place, so changing a quote means canceling the existing order and submitting a replacement”⁶. Kacho lost “several hundred bucks” to bugs the AI wrote and he ran without reading, one of which backed the wrong team⁵⁸. A software firm that sells bot-building says of one client build that most strategy ideas died in testing on past data; it publishes no profit figure⁷³. Nothing we read supports the version where a non-programmer follows a tutorial tweet.
6. Time, and acceptance that an edge fades. The largest accounts took months to grow⁶¹. A typical machine-run wallet in Bitquery’s count ran for 4 months and traded 2,925 markets⁶². Kacho’s best months were his first two⁵⁷.

How crowded it is

Machine-run wallets’ share of the money traded on Polymarket roughly doubled, “from about 16% in September to about 29% in July”⁶². In markets that settle within 15 minutes, the kind the tweets promote, they are on 55.3% of the money⁶². Visible rewards are competed for fast: when about $70,000 of sponsored rewards was added to one market in February 2026, Telonex found makers “saw the new incentive and started competing to post size” within minutes⁷².

The conditions that have to hold

1. You live where a Polymarket site accepts your orders. The main site bars the UK, France, Germany, Italy, Australia and the others above⁸⁹. US residents have Polymarket US, which this page did not study⁵².
2. Either the bot is a maker, or its gain per trade is clearly larger than the taker fee at the prices it trades. In the Telonex week the top wallet was “near 50%” maker, and profitable wallets appeared at every mix¹⁹. Bots that buy late at 90 to 99 cents pay a small fee but risk a large loss on each wrong bet². Being a maker is not protection by itself: the worst result in Bitquery’s August week was a maker’s⁶³.
3. You have money you can afford to lose. No source gives a minimum. Kacho says he started with about $1,000⁵⁸, the 2024 maker with $10,000⁶⁷, and the six-figure accounts bet $2.4 million to $18.2 million in total, though the data does not show their deposits⁶¹.
4. You treat the 5-minute Bitcoin markets with care. Researchers found profit there taken “mostly from retail” before the August 2026 rule change³⁸⁴⁹. In one August 2026 week after fees, 31% of wallets in the 5-minute and 15-minute crypto markets as a whole finished ahead⁶³.
5. You expect the rules to change. They changed many times in 2026¹²⁴⁹.

What it costs

• Tools: tens to a few hundred dollars a month (our estimate). That is a rented server at $4 to $48 a month, and, if you buy them, past market data at $99 to $199, on-chain data at $29 to $199 and an odds feed at up to $249⁷¹. AI coding costs are in Step 7⁴⁵. Tools are not the barrier.
• Trading money: no published figure. Kacho says he started with about $1,000⁵⁸. The 2024 maker says he started with $10,000⁶⁷. The six-figure accounts bet millions, but the data does not show their deposits⁶¹.
• The top speed tier: by approval only. Polymarket’s main servers are in a London data centre, in a country where trading is barred, so ordinary users are pointed to one in Ireland⁸. Placing machines in the London region is offered to those who pass an identity check; no price is published⁸.
• Losses and fees. These remain the main cost. The account “dangerously-skip-permissions” paid about 64% of its profit before fees in fees (our sum)⁶¹.

How to tell early which side you are on

These checks are our judgement. The last four come from Kacho’s account of what went wrong⁵⁷⁵⁸.

• Is your country on Polymarket’s restricted list? Then the main site is closed to you⁸⁹.
• Is your bot a taker? Work out the fee with the formula in Step 2. If your expected gain per trade is smaller, stop.
• Does it still profit with real money and small stakes? Results on past data did not carry over in two of the four losing write-ups³⁰³².
• Are you counting profit after fees? Polymarket’s profile page shows it before fees⁵⁰.
• Is the result from many trades or a few? A high return after a handful of bets is more likely luck; about 60% of “lucky winners” reversed in one study²⁵.
• Must you paste your private key into code you did not write, or pay before you can test? The first has exposed wallet keys³⁴. The second pays the seller whatever happens to you.
• Does the source of the tip gain if you sign up? Look for a tag after a ”?” in the link.
• Are your maker offers being taken, and when? A falling share, or being taken only after the price has moved against you, means faster bots are ahead of you⁵⁸.
• Split your profit into covered and uncovered trades. If the uncovered ones lose, the bot is taking risk it was not built for⁵⁷.
• Is each month smaller than the last? His edge shrank within three months⁵⁷.
• Have you read the code the AI wrote? His unread bugs cost real money⁵⁸.

What we did not verify

• Our first search was cut short. The web-search allowance for the first run ran out early. A second pass on 7 October 2026 searched for success stories, studies since the fees, press coverage and costs. It still found no debunk or confirmation of the student story and no study of copy-trading results.
• Exact quotes. Some pages were read through a tool that extracts or summarises text, so some quotes may not be word-for-word. Only the abstracts of papers³⁵,³⁶,³⁷,³⁸ and³⁹ were read.
• The tweets’ claims. No tweet was opened on X. Our stored copies are cut off at the link, so we did not see images or linked articles. For the student story we checked the linked account’s profit, not its starting money, owner or code. The other unnamed-person stories are unchecked.
• What Polymarket’s data fields mean. They are undocumented. We read “trade_pnl” as profit before fees, “economic_pnl” as what was kept after fees and rebates, and “volume_usdc” as dollars bet, because the sums fit. “Kept” includes open bets at today’s prices. We do not know why the Lucerys profile figure is so high. The leaderboard shows crypto-market profit while the per-wallet figures cover the whole account, so the two are not like for like.
• Who runs any account. For every wallet on this page we do not know the operator, the starting money or the deposits, whether it is a bot, or whether an LLM wrote its code. The data returns nothing for deposits and withdrawals⁶¹.
• Whether “new” wallets are new people, and whether stopped wallets quit. An experienced operator can open a fresh wallet, and a wallet that stops may have been replaced by another.
• The small operator’s own statements. Polymarket’s data confirms the b00k13 wallet’s profit, fees, volume and joining date. His identity, his job and degrees, his starting money of about $1,000, his month-by-month and trade-by-trade figures, and his statement that AI wrote the code are his own. We did not confirm what produced the wallet’s gains since June 2026.
• How many bot operators profit over a long period. We now have one figure for nine days in August 2026, from one firm, whose queries we did not rerun⁶². No source gives the share of bots in profit over months, or a typical bot’s result. Galaxy gives no count of automated accounts in profit⁶⁴.
• Bitquery’s fee rates. Our two readings of its short-markets study gave different figures for fees as a share of volume before and after May 2026, so the page gives none. We did not record that study’s exact dates, and neither Bitquery page was read in full.
• The largest accounts’ monthly figures. Monthly gains for swisstony, RN1 and others are our reading of the profile profit series, before fees.
• Reports we read second-hand. Bloomberg’s analysis is behind a paywall; only its opening was read, and the count of about 800 accounts comes from a news summary⁶⁶. The London Business School and Yale paper was not opened; we read Yale’s write-up⁶⁵ and a news brief²⁵. DeFi Oasis’s data was not opened. Galaxy Research’s report has now been read at source⁶⁴.
• Pew’s sample. It was drawn from traders on 10 busy events, not all wallets, and its method appendix was not opened⁶⁸.
• Press stories. The “$1 into $3.3 million” article was read through a reprint⁶⁹. Starting sums of $1, $63 and $313 in press stories are unchecked. The claim that maker bots “lost a year of profits overnight” in the XRP case rests on a data account on X⁷⁰.
• The 2024 maker’s earnings. They are his own figures in Polymarket’s newsletter, with no wallet given⁶⁷.
• Professional trading firms. Reports that large trading firms are taking stakes in prediction markets in return for market making were seen only as search results and are not used.
• Pages that refused us. The Defiant, bonus.com and help page¹¹ were read through a reader service (r.jina.ai), as were sources⁶² to⁶⁶,⁶⁸,⁷⁰,⁷² and⁷³. A Dune blog post, a Medium article, two write-ups of losing bots, a Korean news site and Kalshi’s fee page did not open and were not used. The last-revision date of help article¹² was not confirmed.
• Copy trading. No study of followers’ results was found. Kreo’s page does not say whether its fee comes on top of Polymarket’s.
• The tagged links. We did not confirm that “?r=” and “?via=” are referral tags that pay.
• Other platforms. Kalshi was not covered. For Polymarket US we read only the fee page and the list of API pages, and did not look for success stories or reward programmes there.
• Costs not priced: the money a competitive market-making bot needs, the price of placing machines near Polymarket’s servers, and tax. Rebates for a small maker rest on one wallet⁶¹. The monthly tool total is our estimate from listed prices⁷¹.
• Enforcement. Polymarket publishes no count of accounts restricted. Regulators outside the US were not searched.
• Whether the people named have seen this page. They were not contacted.

Sources

“Read at source” means we opened the page itself on 7 October 2026, in most cases through a tool that extracts text. “Secondary” means we read an article or summary about the source, not the source itself. “Estimate” means our own sum.

1. Does It Pay collection of 59 tweets on this scheme, collected 2026-10-06; individual tweets linked above. Read from our collected copy on 2026-10-07, not from the live tweets; short links were followed on 2026-10-07. Counts by label and growth rates are estimates.
2. Polymarket documentation, “Fees”. https://docs.polymarket.com/trading/fees . Undated page. Read at source on 2026-10-07. Same figures on the help centre, https://help.polymarket.com/en/articles/13364478-trading-fees , last updated 2026-07-10. Percentages of stake are our estimate.
3. Polymarket documentation, “Maker rebates”. https://docs.polymarket.com/programs/maker-rebates.md and https://docs.polymarket.com/market-makers/maker-rebates . Undated pages. Read at source on 2026-10-07.
4. Polymarket documentation, “Taker rebates”. https://docs.polymarket.com/programs/taker-rebates.md . Programme dated 2026-05-28. Read at source on 2026-10-07.
5. Polymarket documentation, “Liquidity rewards”. https://docs.polymarket.com/programs/liquidity-rewards.md . Undated page; describes an August 2026 programme as ended. Read at source on 2026-10-07.
6. Polymarket documentation, “Market making”. https://docs.polymarket.com/trading/market-making.md . Undated page. Read at source on 2026-10-07.
7. Polymarket documentation, “Rate limits”. https://docs.polymarket.com/api-reference/rate-limits . Undated page. Read at source on 2026-10-07.
8. Polymarket documentation, “Geographic restrictions”. https://docs.polymarket.com/api-reference/geoblock.md . Undated page. Read at source on 2026-10-07 as raw text; counts are ours.
9. Polymarket Terms of Use and page footer. https://polymarket.com/tos . Effective 2026-08-11. Read at source on 2026-10-07 as raw text.
10. Polymarket Help Centre, “Geographic restrictions”. https://help.polymarket.com/en/articles/13364163-geographic-restrictions . Last updated 2026-08-14. Read at source on 2026-10-07.
11. Polymarket Help Centre, “Does Polymarket have trading limits?”. https://help.polymarket.com/en/articles/13364481-does-polymarket-have-trading-limits . 2026-01-11. Read at source on 2026-10-07 through r.jina.ai.
12. Polymarket Help Centre, “Polymarket exchange upgrade: April 28, 2026”. https://help.polymarket.com/en/articles/14762452-polymarket-exchange-upgrade-april-28-2026 . Written for the 2026-04-28 upgrade; date of last revision not confirmed. Read at source on 2026-10-07.
13. Polymarket documentation, “Referral program”. https://docs.polymarket.com/programs/referral-program.md . Terms effective 2026-05-28. Read at source on 2026-10-07.
14. Polymarket documentation, builder fees and tiers. https://docs.polymarket.com/programs/builders/fees.md and https://docs.polymarket.com/programs/builders/tiers.md . Undated pages. Read at source on 2026-10-07.
15. Polymarket public leaderboard data. https://data-api.polymarket.com/v1/leaderboard (crypto, last month, by profit and by volume; overall, all time and last month, by user name). Live data. Read at source on 2026-10-07 at about 11:44 UTC. Medians and shares of accounts are our estimates.
16. Polymarket transparency and integrity pages. https://polymarket.com/transparency and https://integrity.polymarket.com/ . Undated pages. Read at source on 2026-10-07.
17. Polymarket, “agents” code repository. https://github.com/Polymarket/agents . Last changed 2024-11-05. Read at source on 2026-10-07.
18. Saguillo, Ghafouri, Kiffer and Suarez-Tangil, study of arbitrage on Polymarket, arXiv 2508.03474. https://arxiv.org/abs/2508.03474 (full text https://arxiv.org/html/2508.03474v1 ). 2025-08-05. Read at source on 2026-10-07. Monthly, per-transaction and top-ten figures are our estimates.
19. Telonex, “Top crypto traders on Polymarket 15-minute markets”. https://telonex.io/research/top-crypto-traders-polymarket-15m . 2026-02-13. Read at source on 2026-10-07. Telonex sells market data. Monthly pool is our estimate.
20. Andrey Sergeenkov, Polymarket profitability report. https://sergeenkov.com/polymarket-profitability/ . 2026-04-06. Read at source on 2026-10-07; secondary for DeFi Oasis’s data.
21. The Defiant, report of²⁰. https://thedefiant.io/news/research-and-opinion/polymarket-profitability-report-april-2026 . 2026-04-06. Secondary; read through r.jina.ai.
22. Bonus.com, report of Galaxy Research’s analysis. https://www.bonus.com/news/galaxy-research-nearly-7-in-10-retail-polymarket-traders-ended-up-losing-money/ . 2026-10-02. Secondary; read through r.jina.ai.
23. Finance Magnates, report of a Bloomberg analysis. https://www.financemagnates.com/fintech/100000-polymarket-wallets-lost-at-least-1000-bloomberg-analysis-shows/ . 2026-04-29. Secondary.
24. PYMNTS, report of the same Bloomberg analysis. https://www.pymnts.com/markets/2026/most-polymarket-users-lose-money-as-top-accounts-win-big . 2026-04-28. Secondary.
25. KuCoin news brief on a London Business School and Yale study. https://www.kucoin.com/news/flash/3-14-of-polymarket-traders-account-for-30-of-profits-67-of-users-bear-all-losses . 2026-04-27. Secondary.
26. Decrypt, “Most Polymarket traders have lost money”. https://decrypt.co/290625/most-polymarket-traders-have-lost-money . 2024-11-07. Secondary.
27. Tangem news, on Polymarket profits. https://tangem.com/en/news/markets/7806-polymarket-profits-70-lose-tiny-elite-take-3-7b/ . 2025-12-29. Secondary.
28. DEV Community post by a bot operator, “February 2026 changed Polymarket forever”. https://dev.to/lkto1m/february-2026-changed-polymarket-forever-heres-what-happened-to-my-bots-numbers-2fi5 . 2026-06-25. Read at source on 2026-10-07 for the author’s own figures; secondary for Polymarket’s rule changes. Daily and monthly sums and trade counts are our estimate.
29. Cointelegraph, “Polymarket quietly adds taker fees to 15-minute crypto markets”. https://cointelegraph.com/news/polymarket-quietly-adds-taker-fees-15-minute-crypto-markets . 2026-01-06. Secondary.
30. Falsify Lab, “Strategy deep dive: Polymarket”. https://falsifylab.substack.com/p/strategy-deep-dive-polymarket . 2026-06-27. Read at source on 2026-10-07.
31. Startup Fortune, “I actually gave an AI money to trade on Polymarket”. https://startupfortune.com/i-actually-gave-an-ai-money-to-trade-on-polymarket-heres-what-nobody-on-twitter-wants-you-to-know/ . 2026-02-12. Read at source on 2026-10-07.
32. Purpose Investments, “What I learned losing money to AI in a prediction market”. https://www.purposeinvest.com/thoughtful/what-i-learned-losing-money-to-ai-in-a-prediction-market . 2026-05-07. Read at source on 2026-10-07.
33. Filthy Monkey newsletter, “I gave an AI $100 and told it to make money”. https://filthymonkey.beehiiv.com/p/i-gave-an-ai-100-and-told-it-to-make-money . 2026-02-19. Read at source on 2026-10-07.
34. StepSecurity, report on Polymarket bot repositories that take wallet keys. https://www.stepsecurity.io/blog/malicious-polymarket-bot-hides-in-hijacked-dev-protocol-github-org-and-steals-wallet-keys . 2026-03-15. Read at source on 2026-10-07.
35. Gregory Young, arXiv 2607.26245, a Polymarket and Binance dataset for 15-minute Bitcoin markets. https://arxiv.org/abs/2607.26245 . 2026-07-28. Read at source on 2026-10-07 (abstract only).
36. arXiv 2604.14199, live benchmark of language models on Polymarket. https://arxiv.org/abs/2604.14199 . 2026-04-03. Read at source on 2026-10-07 (abstract only).
37. arXiv 2608.00666, study of arbitrage in negative-risk markets. https://arxiv.org/abs/2608.00666 . 2026-08-01. Read at source on 2026-10-07 (abstract only).
38. arXiv 2606.31675, study of manipulation in short Bitcoin markets. https://arxiv.org/abs/2606.31675 . 2026-06-30. Read at source on 2026-10-07 (abstract only).
39. arXiv 2606.16852, study of orders made to fail at settlement. https://arxiv.org/abs/2606.16852 . 2026-06-15. Read at source on 2026-10-07 (abstract only).
40. Moon Dev sales page. https://moondev.com/go . Undated page, copyright 2026. Read at source on 2026-10-07.
41. Kreo documentation: index, fees, referral rewards, copy-trade rewards, terms of service. https://docs.kreo.app/index.md , https://docs.kreo.app/fees.md , https://docs.kreo.app/Referral-Rewards.md , https://docs.kreo.app/copy-trade-rewards.md , https://docs.kreo.app/terms-of-service.md (terms updated 2026-03-15; others undated). Read at source on 2026-10-07.
42. PolyTrack pricing. https://www.polytrackhq.app/pricing . Undated page. Read at source on 2026-10-07 from the site’s script file, not the displayed page.
43. DefiLlama, Polymarket fees and revenue. https://api.llama.fi/summary/fees/polymarket?dataType=dailyFees (and dataType=dailyRevenue). Live data, 2026-10-07. Secondary (a third party’s count).
44. poly-maker code repository. https://github.com/warproxxx/poly-maker . Last code change 2026-07-09. Read at source on 2026-10-07.
45. Anthropic, Claude pricing. https://platform.claude.com/docs/en/about-claude/pricing . Undated page. Read at source on 2026-10-07.
46. US Commodity Futures Trading Commission, customer advisory “AI Won’t Turn Trading Bots into Money Machines”. https://www.cftc.gov/LearnAndProtect/AdvisoriesAndArticles/AITradingBots.html . 2024-01-25 (date from release 8854-24, https://www.cftc.gov/PressRoom/PressReleases/8854-24 ). Read at source on 2026-10-07.
47. US Commodity Futures Trading Commission, release 8478-22, “CFTC Orders Event-Based Binary Options Markets Operator to Pay $1.4 Million Penalty”. https://www.cftc.gov/PressRoom/PressReleases/8478-22 . 2022-01-03. Read at source on 2026-10-07.
48. US Commodity Futures Trading Commission, prediction markets advisory, release 9185-26. https://www.cftc.gov/PressRoom/PressReleases/9185-26 . 2026-02-25. Read at source on 2026-10-07.
49. Polymarket documentation, “Changelog”. https://docs.polymarket.com/changelog.md . Dated entries from 2026-01-05 to 2026-09-29. Read at source on 2026-10-07 as raw text.
50. Polymarket per-wallet data: statistics at https://data-api.polymarket.com/v2/user-stats?user= and the profile profit series at https://user-pnl-api.polymarket.com/user-pnl?user_address=&interval=all&fidelity=1d ; field notes at https://docs.polymarket.com/trading/wallet-activity.md . Live data. Read at source on 2026-10-07 at about 11:44 UTC. Shares, percentages and the meaning of undocumented fields are our estimates.
51. Polymarket documentation, “Trading rate limits”. https://docs.polymarket.com/api-reference/trading-rate-limits.md . Undated page; refers to 2026-07-24. Read at source on 2026-10-07.
52. Polymarket US documentation, “Fee schedule” and page index. https://docs.polymarket.us/fees.md and https://docs.polymarket.us/llms.txt . Fees effective 2026-10-01. Read at source on 2026-10-07.
53. US Commodity Futures Trading Commission, release 9217-26 (complaint, pending when read, against a US service member). https://www.cftc.gov/PressRoom/PressReleases/9217-26 . 2026-04-23. Read at source on 2026-10-07.
54. US Commodity Futures Trading Commission, release 9237-26 (complaint, pending when read, against an employee of a technology company). https://www.cftc.gov/PressRoom/PressReleases/9237-26 . 2026-05-27. Read at source on 2026-10-07.
55. US Commodity Futures Trading Commission, release 9276-26 (order with a three-year trading ban; no platform named). https://www.cftc.gov/PressRoom/PressReleases/9276-26 . 2026-07-31. Read at source on 2026-10-07.
56. US Commodity Futures Trading Commission, release 9289-26 (order with a three-year trading ban; Kalshi assisted). https://www.cftc.gov/PressRoom/PressReleases/9289-26 . 2026-08-28. Read at source on 2026-10-07.
57. Kacho, “Polymarket arbitrage: real numbers” (a bot operator’s own account, naming the wallet b00k13). https://kacho.io/polymarket-arbitrage-real-numbers . 2026-06-06. Read at source on 2026-10-07. The author’s own figures; the wallet total is checked in⁶¹.
58. Kacho, “Why my Polymarket arbitrage bot lost money”. https://kacho.io/why-my-polymarket-arbitrage-bot-lost-money . 2026-06-30. Read at source on 2026-10-07. The author’s own figures.
59. polymm code repository (the same operator’s free bot code). https://github.com/kachence/polymm . Created 2026-07-19, last changed 2026-08-16. Read at source on 2026-10-07.
60. Polymarket public leaderboard data, crypto category, top 50 by profit, all time and last month. https://data-api.polymarket.com/v1/leaderboard?category=CRYPTO&timePeriod=ALL&orderBy=PNL&limit=50 and https://data-api.polymarket.com/v1/leaderboard?category=CRYPTO&timePeriod=MONTH&orderBy=PNL&limit=50 . Live data. Read at source on 2026-10-07 at about 14:00 UTC. Counts of accounts are our estimates.
61. Polymarket per-wallet data, second reading, for the wallets named under “The upside” and the 100 wallets in⁶⁰: the same two addresses as⁵⁰, for example https://data-api.polymarket.com/v2/user-stats?user=0x1c5575dc20e4ea54d1bb09ccda72ccf8a3b684ce (b00k13), https://data-api.polymarket.com/v2/user-stats?user=0xc53375ff94e96100f2b30a4b5775db35218d69a9 (dangerously-skip-permissions), https://data-api.polymarket.com/v2/user-stats?user=0xe9076a87c5ed90ef16e6fe6529c943baeca0cff6 (suntori) and https://user-pnl-api.polymarket.com/user-pnl?user_address=0x204f72f35326db932158cba6adff0b9a1da95e14&interval=all&fidelity=1d . Live data. Read at source on 2026-10-07 at about 14:00 UTC. Shares, monthly gains, counts and the meaning of undocumented fields are our estimates.
62. Bitquery, “How much of Polymarket is bots”. https://bitquery.io/investigations/how-much-of-polymarket-is-bots . 2026-08-20. Read at source on 2026-10-07 through r.jina.ai. Bitquery sells blockchain data.
63. Bitquery, study of Polymarket’s 5-minute and 15-minute crypto markets. https://bitquery.io/investigations/polymarket-five-minute-casino . 2026-08-19. Read at source on 2026-10-07 through r.jina.ai. Bitquery sells blockchain data.
64. Galaxy Research, on-chain analysis of Polymarket trader behaviour. https://www.galaxy.com/insights/research/polymarket-trader-behavior-onchain-analysis-retail-profit-loss-specialization . 2026-10-01. Read at source on 2026-10-07 through r.jina.ai; charts not read. Galaxy is a crypto financial firm.
65. Yale Insights, “Wisdom of the few: prediction markets are driven by a small number of skilled traders” (the university’s write-up of the London Business School and Yale study). https://insights.som.yale.edu/insights/wisdom-of-the-few-prediction-markets-are-driven-by-small-number-of-skilled-traders . 2026-06-10. Secondary; read through r.jina.ai.
66. Newser, “Polymarket’s biggest winners aren’t human” (summary of the Bloomberg analysis of 2026-04-28). https://www.newser.com/story/388532/polymarkets-biggest-winners-arent-human.html . 2026-05-10. Secondary.
67. Polymarket newsletter, “Automated market making on Polymarket” (interview with the author of⁴⁴). https://news.polymarket.com/p/automated-market-making-on-polymarket . 2025-05-28. Read at source on 2026-10-07. Earnings are the interviewee’s own figures.
68. Pew Research Center, study of 11,989 active Polymarket wallets. https://www.pewresearch.org/?p=343505 . 2026-07-22. Read at source on 2026-10-07 through r.jina.ai; method appendix not opened.
69. KuCoin news brief reprinting a Finbold article, “AI-powered trading bot turns $1 into $3.3M on Polymarket”. https://www.kucoin.com/news/flash/ai-powered-trading-bot-turns-1-into-3-3m-on-polymarket . 2026-04-03. Secondary.
70. CoinDesk, carried by Yahoo Finance, “Polymarket trader nets $233,000…”. https://finance.yahoo.com/news/polymarket-trader-nets-233-000-061703636.html . 2026-01-19. Secondary; read through r.jina.ai.
71. Vendors’ price pages: Telonex https://telonex.io/pricing , Bitquery https://bitquery.io/pricing , The Odds API https://the-odds-api.com/ , Amazon Lightsail https://aws.amazon.com/lightsail/pricing/ , DigitalOcean https://www.digitalocean.com/pricing/droplets . Undated pages. Read at source on 2026-10-07 through r.jina.ai. The monthly total is our estimate.
72. Telonex, study of sponsored rewards in one Polymarket market. https://telonex.io/research/sponsored-liquidity-rewards-jesus-market . 2026-02-22. Read at source on 2026-10-07 through r.jina.ai. Telonex sells market data.
73. Valletta Software, blog post on building a Polymarket trading bot with Claude. https://vallettasoftware.com/blog/post/polymarket-trading-bot-claude-fable-5 . 2026-07-22. Read at source on 2026-10-07 through r.jina.ai. The firm sells bot-building services.
74. Kacho, “About”. https://kacho.io/about . Undated page. Read at source on 2026-10-07. The author’s own statement.

Corrections

If you are quoted or named on this page and think something is wrong, want your reply shown beside it, or want to be removed, write to [email protected]. We aim to reply within 14 days, and always within one month. Factual errors are corrected with a dated note here. A sentence that is seriously disputed comes down while we check it.

Change log:

• 7 October 2026: second research pass. Added “The upside” and “What it takes to compete” in place of “When it can work”. Corrected the heading “Bots as a group win” after a count made after the current fees found bot-run wallets behind as a group⁶². Corrected the statement that no checked success by an individual with an AI-built bot had been found⁵⁷⁶¹. Added a study of results since the current fees⁶⁸. Sources 57 to 73 added.
• 7 October 2026: review of the second pass. Corrected the source for the small operator’s starting money to⁵⁸ and for his degrees to a new source⁷⁴. Added his own caution that he cannot separate faster bots from his move into sports markets⁵⁸. Corrected RN1’s monthly range to exclude September 2026, and the trade counts of three new accounts⁶¹. Made clear that the 31% figure covers the 5-minute and 15-minute crypto markets together⁶³.
• 2026-10-07: wording about named accounts and companies reviewed; the CFTC cases are now given as pending allegations or settled orders, link tags that pointed to two accounts are no longer printed, one example tweet was replaced, and statements about what tweets leave out now say that only our cut-off copies were read.