Racing the oracle: Polymarket's 26-second arbitrage
A Polymarket crypto market ends at a time printed in its own title, but the chain is not told the result for another half minute or so. We read 551,453 of these markets, worked out the exact second each one closed, and followed every bet placed after the clock ran out and before the answer arrived.
At five past seven in the morning, New York time, on 3 May 2026, a 5-minute bet on the price of Solana ran out of time. It had been a quiet little market. People had put $570 into it, in bets averaging a couple of dollars, and most of that money had backed the price to finish higher. The side betting it would finish lower was changing hands at about 37 cents on the dollar. Then the clock stopped.
Nothing was settled yet. Polymarket is a site where people bet on how real events turn out, and a market like this one is decided by the price of Solana at the moment the clock stops. That price comes from a feed outside the site, and it is not written onto the public record until a separate transaction lands a little later. Here that took 36 seconds, and for those 36 seconds the market stayed open for business while the answer already existed somewhere.
Half a million dollars was spent in that half minute, something like 909 times everything the market had taken while it was actually running. All of it went one way, onto the side the crowd had written off, and that side won.
Knowing the answer slightly before the person on the other side of the deal is the oldest edge in trading. Bitquery decodes every Polymarket trade on Polygon, so we went looking for how often it happens and what it is worth. We read 551,453 of these short crypto markets, every one settled between September 2025 and this August. From each market's own title we worked out the exact second its clock ran out, and compared that against the moment the chain was told the result. Then we measured what every bet placed in between actually returned. Buying the written-off side in that gap returned 1,030% across the whole record, and almost all of it was won in 3 weeks.
01 — The premiseWhy there is a gap at all
The bets on Polymarket trade like shares. If you think a thing will happen you buy the "yes" at whatever the crowd is charging, somewhere between nothing and a dollar. If it happens you get a dollar a share, and if it does not you get nothing. The price is therefore a probability wearing a dollar sign, and 97 cents means the crowd is almost sure.
Most of the trading there now is on a much smaller question, whether a coin will be worth more in a few minutes than it is right now. Bitcoin, Ethereum, Solana and 5 others run on a loop, in three window lengths, and the shortest closes every 5 minutes around the clock. The title of each one prints its window in New York time, so a market reads "Solana Up or Down - May 3, 7:00AM-7:05AM ET" and anybody can see when it ends. We counted these markets in an earlier piece on the 5-minute casino, which is now most of what the venue does.
Ending is not the same as being settled. The result comes from a price feed run by Chainlink that lives outside the chain. Polymarket reads it, decides which side won, and only then tells the chain. The number itself is never written down anywhere public. What lands on the chain is only the verdict, and it lands after a delay. Across the record the middle of that delay is 26 seconds, and nine markets in ten are settled within 85 seconds of the clock stopping.
During that gap the order book stays open. Anyone can buy and anyone can sell, at whatever price the book will bear. The difference is that the thing being bet on has already happened.
02 — The sceneOne market, 36 seconds
Go back to the Solana market. Its whole life is in the record. It took small bets for 5 minutes, and the side backing a fall drifted down to around 37 cents, which is the crowd saying it probably will not happen. Then the clock stopped and half a million dollars arrived.
The buying came from 184 wallets and it swept the book clean. One of them took every offer on the way up within a single second, starting at 27 cents and finishing at 98. That is not somebody forming a view about Solana. That is somebody who already knows, buying whatever is still cheap before the rest of the market catches up.
The cheapest of it went for about a penny a share. Mep72 spent $46 and got back nearly sixty times that. Larger and slower money came in behind it at prices close to a dollar, which is a duller version of the same trade and a much safer one.
We checked the market against the outside world. Solana on Binance drifted down across the window, by a few hundredths of a percent. The market settled the way the tape did.
03 — The measurementWhat the gap is worth
One market proves nothing. So we did the same arithmetic on all 551,453 of them, splitting every bet by when it was placed: while the market was live, in the gap, or after the chain had already published the answer. Then we asked the only question that matters about a bet, which is what it paid.
| When the money was spent | On the near-certain side (99c and up) | On the written-off side (3c and under) | ||||
|---|---|---|---|---|---|---|
| Spent | Result | Return | Spent | Result | Return | |
| Before the clock stopped | $373,411,771 | −$2,869,951 | −0.8% | $13,179,696 | +$12,395,507 | +94.1% |
| In the gap, before the chain was told | $898,552,079 | +$373,002 | +0.0% | $2,240,651 | +$23,083,874 | +1,030.2% |
| After the chain was told | $175,099,085 | −$56,645 | −0.0% | $9,716 | −$3,596 | −37.0% |
Trading almost stops once the result is public, which makes sense, because by then everybody can see it. The written-off side is the row to look at. Bought during the gap it returned 1,030%, which is roughly $2.2M going in and more than ten times that coming back.
Buying at 99 cents or better is a dull trade by design, and across the whole record it came out roughly flat. It does not stay dull as the clock runs down.
The price these buyers paid moves by a hundredth of a cent across the three windows, while what they were actually buying got steadily worse. It came home 99.15% of the time earlier in the market and 97.75% of the time once the clock had stopped. A book can only reprice on what it can see, and it cannot see the feed that decides the market.
04 — The three weeksWhen it stopped being a trickle
Spread across a year, $23.1M sounds like a steady tax on the slow. It was not steady at all. For most of the record, penny bets placed in the gap won about as often as penny bets anywhere else, which is to say almost never. Then, in the middle of April 2026, they started winning.
| Week beginning | Bets placed | Money spent | Share that won | Result |
|---|---|---|---|---|
2026-04-05 | 34,644 | $20,143 | 3.34% | +$29,613 |
2026-04-12 | 32,828 | $20,874 | 7.98% | +$104,198 |
2026-04-19 | 114,866 | $53,560 | 25.94% | +$995,865 |
2026-04-26 | 357,943 | $244,012 | 67.13% | +$11,706,973 |
2026-05-03 | 112,841 | $73,623 | 84.91% | +$4,628,820 |
2026-05-10 | 12,798 | $4,451 | 3.44% | +$7,645 |
2026-05-17 | 8,476 | $3,665 | 2.22% | +$3,149 |
2026-05-24 | 3,973 | $2,141 | 0.19% | −$1,797 |
2026-05-31 | 2,432 | $1,225 | 0.00% | −$1,225 |
These are bets priced at a bit over a cent, which is the market calling them hopeless. Across the 3 weeks from the middle of April to the second week of May, more than half of the pennies staked in the gap ended up on the winning side, and at the peak more than 4 of every 5 did. Those weeks produced $17.3M. Then it stopped as abruptly as it had started, and it has not come back.
The obvious explanation is that the market was wild and prices were whipping around. The opposite is true. We pulled Bitcoin's own 5-minute moves from Binance for every month in the record, and May 2026 was one of the calmest months in the whole record. February was more than twice as jumpy and produced almost none. Whatever opened the door in April, it was not volatility.
We can say what changed on the chain and we cannot say why. What the record shows is that the supply of cheap resting offers sitting in the gap dried up in the second week of May, and the winning stopped at the same moment. Whether somebody at Polymarket changed how long orders survive the closing bell, or the market makers on the other side simply learned to pull their quotes, is not something a ledger records.
05 — The two sidesWho pays for it
Somebody is on the other end of all this. On a matched trade the record names both parties, so we can look at the buyers of the near-certain side in the gap and at the people selling to them. They are not the same kind of crowd. In August the selling came from 12,403 wallets and the buying from 1,188, and the small side is the one holding the losses.
A fair objection is that this is one leg of a bigger book. A market maker who loses a little on every fill can still finish the day ahead, and judging a wallet on a single kind of trade is how people get this wrong. So we took those buyers and added up everything they did, across every market that opened and settled inside the same window.
| Wallets | Result on everything | Fees paid | After fees | Of which, the gap | |
|---|---|---|---|---|---|
| Bought the near-certain side in the gap | 1,157 | −$282,313 | $678,926 | −$961,239 | −$238,453 |
| Everyone else | 60,166 | +$568,258 | $3,546,504 | −$2,978,246 | — |
It does not come back. The gap is where their losses sit, the rest of their trading is close to a wash, and they pay fees on top. Their median wallet finishes $113 down over 3 weeks, against $7 for everybody else. Many of them are automated: our count of Polymarket's machines found the same pattern, that the round-the-clock wallets were behind before fees as well as after.
06 — The winnersA small, busy crowd
Close to 8,000 wallets tried the penny side of the gap during April and May, and about 4 in 10 of them came out ahead, which sounds like a broad opportunity a lot of people noticed. The money did not spread evenly across them at all.
| Wallet | Bets | Markets | Spent | Won | Multiple | Account |
|---|---|---|---|---|---|---|
Mep72 | 65,257 | 5,465 | $93,729 | +$6,874,804 | 73× | opened 2025-12-21 |
95slave | 2,999 | 567 | $20,645 | +$1,208,597 | 59× | opened 2026-04-28 |
0xd02c…b1d7 | 9,596 | 880 | $13,375 | +$907,168 | 68× | opened 2025-12-21 |
0x619e…f583 | 4,346 | 750 | $9,897 | +$744,814 | 75× | opened 2025-12-15 |
0x7fa1…347d | 2,540 | 846 | $12,021 | +$730,194 | 61× | no profile |
PayoRamon | 975 | 198 | $10,565 | +$646,363 | 61× | opened 2026-04-28 |
NotWinr | 5,485 | 3,611 | $7,070 | +$610,531 | 86× | opened 2026-04-28 |
egig | 7,328 | 4,156 | $22,629 | +$532,639 | 24× | opened 2026-03-07 |
The largest single wallet took 30.5% of the total on its own, and the top 10 took most of the rest. That is concentrated enough to be worth saying plainly: this was a small group, not a crowd. Running every one of them against the round-the-clock test, 80.4% of the money went to wallets that place orders in every hour of the day, rarely more than a minute apart. Three of the biggest winners opened their Polymarket accounts on the very same day, in the week the trade was at its best.
What none of that tells you is who they are. A profile name is a name somebody typed. Trading in the same second as somebody else is not proof of being the same somebody. Speed is the only thing the record demonstrates here.
07 — The correctionA sure thing is safer than we thought
Our own earlier reading of this data suggested that near-certain bets on a rise failed about 5.7% of the time while near-certain bets on a fall almost never did, a gap of about 60 times. It does not survive a recount. Measured against the last price set before the clock stopped, the two sides fail at almost the same rate.
| Markets counted | "Up" was near certain | "Down" was near certain | Gap between them |
|---|---|---|---|
| Every market | 1.74% of 200,721 | 2.02% of 202,710 | 0.86× |
| Only markets that traded properly | 1.95% of 111,065 | 1.53% of 109,936 | 1.27× |
"Traded properly" means at least 200 bets were placed before the clock stopped. Thin markets are the ones where a single small bet sets the last price, and they pull the two columns in opposite directions.
The flaw was in which price was used. The earlier number took each market's final price, including the trades that happen in the gap. Once the clock stops, the side that is about to lose gets marked down toward nothing by exactly the fast money described above, so a market about to spring a surprise quietly leaves the near-certain group before anybody counts it. That empties one column and not the other, and manufactures a gap that was never there. The honest figure is that a price of 97 cents or better is wrong a little under 2% of the time, about 1 market in 50, whichever side it favours.
08 — The feeWhat it did not do
Polymarket charges the aggressive side of a trade a fee, introduced on these markets in January 2026 and cut sharply at the start of May. The company's own page gives it a plain purpose, funding rebates for the traders who rest orders, and says nothing about speed. The press read it differently: coverage at launch called it a brake on latency arbitrage, because it bites hardest at even odds, where the fastest money works. If it were holding the line either way, the record would show it.
The timing runs the wrong way. The fee had been in force for 3 months when the gap swallowed half of all the money spent on these markets in May, the high-water mark of the whole record. Fees charged inside the gap during the best week of the episode were lower than they had been in March, and the cut at the start of May made the trade cheaper a week before it died. In the peak week the fee worked out to close to 3 cents on every dollar staked in the gap, set against bets that paid out at better than fifty to one.
The ordinary version of the trade has been ground down anyway. Buying the near-certain side in the gap ran mildly profitable through the spring and has been negative every month since June, which is what competition normally does to an edge that anyone can copy. The spectacular version ended in a single week, and the chain does not record who ended it. Polymarket is now a large share of everything happening on Polygon, which we measured separately.
09 — The recordAddresses and markets
Everything above can be checked against the chain. The market in the opening is condition 0x298617f33f8e7908…, settled in May. Queries against the same data run through the prediction market API and the Polymarket examples. These are the wallets named above.
Mep72— won $6,874,804 in the gap across 5,465 markets in April and May 202695slave— won $1,208,597 in the gap across 567 markets in April and May 20260xd02c…b1d7— won $907,168 in the gap across 880 markets in April and May 20260x619e…f583— won $744,814 in the gap across 750 markets in April and May 20260x7fa1…347d— won $730,194 in the gap across 846 markets in April and May 2026PayoRamon— won $646,363 in the gap across 198 markets in April and May 2026PayoRamon— $26,178 from the Solana market in the openingthe settling contract— reports the verdict for every crypto up or down market
How this was measured
The data. Every Polymarket trade and settlement on Polygon, decoded by Bitquery, from 2025-09-01 to 2026-08-22 (August is a partial month). We kept the 551,453 short crypto markets that print a clock in the title, were settled, and settled to a plain up or down. Markets whose clock falls on a daylight saving changeover are excluded, 20 in total. Markets whose title gives only one time, rather than a window, are excluded too: the printed time is not their closing time.
The clock. Closing times are read from each title in New York time, which handles the summer and winter changeover on its own. The check is that a market must be settled after it closes, never before: that holds for every market in the set once the changeover days are removed.
What a return means here. Every figure is what money spent buying actually came back: shares of the winning side, at a dollar each, less what was paid for them. Prices on these markets sit on whole cents, so the price groups are drawn a hair beyond their stated edges; without that, a bet at exactly 3 cents falls in or out of the group on floating point noise. It is deliberately not a full ledger of winners against losers. For some trades in the gap the aggressive party has no summary line of its own in the decoded record, so the two sides of a single market do not always cancel, and we do not present them as if they do. The one place we do name both sides, in section 05, the two halves agree to within a couple of dollars on $14 million.
What cannot be known. The price that decides these markets is never written on the chain. Polymarket's settling contract publishes the verdict and nothing else. So a surprise is measured against the market's own last price, not against the number that actually settled it, and we cannot say by how much any market was won or lost. A price of 97 cents is the crowd's confidence, not the feed's. Our outside check against Binance reconciles 3 of the 4 surprises we sampled; the one that does not moved 1.32 hundredths of a percent, which is smaller than the difference between a spot price and the sixty second average these markets are settled on. That is the limit of what an outside price can prove.
What speed does not prove. Trading first shows that a wallet was fast. It does not show how it knew, and it does not show who runs it. We name wallets only by the profile they chose on Polymarket.
Coverage. Months before May 2026 were rebuilt from the chain after the fact and run about 95% complete, so shares within a month are sound and absolute totals for those months are floors. August 2026 is a partial month throughout. The first days of May 2026 sit across a change in Polymarket's own contracts and our volumes there are less certain than elsewhere.
Run these queries yourself
Every number here comes from Polymarket's own trades and settlements on Polygon, decoded by Bitquery. The same records are open through the Polymarket API, and Bitquery's MCP server puts them inside Claude and other AI tools, so you can ask which markets kept trading after their clock stopped in plain language.
Connect the MCP server