Field note No.01 · 2026-08-05
How many pump.fun launches actually graduate?
No sampling, no estimates, no borrowed numbers. Data released under CC0.
Every figure you have read about this is probably true, which is exactly the problem. Estimates float from a fifth of a percent to six percent, and almost nobody says what they divided by. So here is one dataset counted six different ways, with the denominator written down each time.
The whole disagreement, in one number
Between the lowest and the highest honest answer. Same tokens, same month, same chain — only the denominator changes.
It depends what you call a launch
A token appears on pump.fun about every three seconds, and most are never touched by anyone. Whether those stay in the denominator moves the answer more than the market ever does.
15% of launches never see a single trade. Created and abandoned inside the same minute — the token equivalent of a domain nobody ever points at anything. Keep them and you get 2.62%. Drop them and you get 2.98%. Neither is a lie; they answer different questions.
If someone quotes you a graduation rate without a denominator, they are quoting a number they did not measure.
The two we would defend: 2.62% of everything minted, and 4.15% of tokens that got as far as one SOL — the second being closer to what a trader actually meets, because a token without a single buy never reaches anybody's screen.
The part we did not expect
The creator's own first buy decides more than anything measurable later
When a token is created its author may buy some of it on the spot. The amount is public, it is fixed before anyone else can react — and it splits outcomes harder than most things you can compute afterwards.
It is not a line, and that is the whole finding. A creator who buys nothing at all lands at 0.34% — about 8 times below average dead on arrival
From there it climbs to 5.70% around 0.05–0.2 SOL, sags through the middle, then jumps again to 7.98% once the creator commits more than five. Two different species sit on those two peaks: the ordinary launch where somebody takes a normal position in their own coin, and the heavily prepared one where the opening buy is large and deliberate. The trough between them — 0.5–1 SOL at 1.45% — is where half-hearted attempts go to die.
If it happens, it happens fast
Median time from creation to graduation. 81% of all graduations are finished inside the first hour.
The slow tail is real — the ninetieth percentile sits at 557 minutes — but a token that has been quiet for an hour has already told you what it is.
And it moves, day to day
Daily graduation rate across the 31 days measured: low of 1.43%, high of 3.62%. Anyone quoting a single decimal place without a window is quoting noise.
We are the outlier — so we went and found out why
Our number is higher than everyone else's
Published estimates for the same question sit far below ours: an academic survey of 832,941 launches reports 0.198%, DEXTools 0.26%, Bitget 1.15%. Our lowest honest answer is 2.62%. Someone here is wrong, and we would rather find out than be quoted.
So we did two things: audited our own flag, then took the one estimate whose author had published her raw data and reproduced it.
First the flag. We read the bonding-curve account straight off the chain for a random sample of one ten-day window: 249 of 249 tokens we mark as graduated carry the completed flag, and 0 of 750 tokens we do not mark carry it. Separately we asked a third-party aggregator whether each token trades on an AMM off the curve — a pool on the curve itself does not count, and that trap is what made our first attempt at this check useless.
The aggregator confirmed 73.3% of the graduated sample against 0.3% of the control. We used to quote that 73.3% as our accuracy. It is not — it is the aggregator's coverage, and the missing quarter are tokens that migrated and died faster than it indexed a live pool. The on-chain read above is the accuracy number, and we should have said so the first time.
Every confirmed migration landed on pump.fun's own AMM — 220 pools — plus a handful on Meteora and one on Orca. Not one on Raydium. Any tool still defining graduation as "a Raydium pool appeared" is counting a road traffic stopped using. That failure mode is real and we have met it. It is not, as it turns out, what explains the academic number.
The correction
We compared across different months, which is the same sin
The 0.198% comes with its dataset attached — 860,213 launches, CC-BY, on Zenodo. So we downloaded it instead of arguing. Two things turned out to be true at once.
The first is that her file measures a shorter clock than its label. Recomputed from the raw timestamps, every graduation in it lands within 5.98 minutes of creation and every timeout at exactly 1440. Across 833,171 records, the span between six minutes and twenty-four hours holds zero outcomes. A stated 24-hour rate is a six-minute rate. Her own file shows the cost: graduated tokens top out at 165 SOL, yet 8,323 timed-out ones finish above 115 SOL and 496 above 1,000 — levels no token reaches while still on the curve.
The second is ours, and we had not accounted for it. Her window closes 2026-06-10; our data begins 2026-06-12. The two never overlap by a single day. Our nearest slice, 12–21 June, graduates at 0.8173% — not 2.62%. Truncate it at her six minutes and it reads 0.2446%, against her 0.207%.
Match the period and match the window, and the two measurements agree to within about twenty percent. Her number is right for what it measures; the label is what misleads. And our own headline, held up against hers with no date range attached, was committing the identical error in the opposite direction. A graduation rate needs a denominator, a venue, and a window. We got the third one wrong on this page, and we would rather write that here than quietly edit the number and hope nobody diffed it.
Ours: every launch we observed, migration read from the on-chain event, window stated above. Check it — the calls dataset and its mirror are public.
How this was measured
Every launch created between 2026-07-06 and 2026-08-05 is included, minus the last six hours so recent tokens had time to resolve. Graduation means the token migrated off the bonding curve, read from the on-chain event rather than inferred from price. Curve progress is measured in SOL raised. The creator's opening buy comes from the creation event itself. No sampling — this is all of them, 825,123.
Worth stating plainly: one thirty-day window, one chain, and a cool market — the daily range above shows how much that alone moves things. And a token that graduates is not a token that made anyone money. This page says nothing about that. Our own call record, losses printed alongside the wins, is on the front page.
Every call, hashed and timestamped on Bitcoin so it cannot be edited afterwards. Schema and licence at /data.html. Take it, republish it, argue with it — no permission needed.