# How we rate

Perp Reference · Method note · Updated: 24 September 2026

Every score on this site comes out of one method. It was fixed on 22 September 2026, before a venue had
been read or a book had been measured, and nothing in it has moved since: not a weight, not a rubric,
not the point at which a curve stops paying. Totals are printed on a scale of 0 to
7, to 1 decimal place, and
[current scores](#current-scores) is where the 10 scored venues stand today.

Six criteria divide a hundred points between them. Each one is worked out on its own rubric to a mark
out of ten, the six marks are combined in the proportions set out below, and the result is put on this
scale. The weights are settled here and nowhere else; where another page prints them, it prints them
from this one. An entry chooses which venues it covers and which number puts them in sequence; it
never chooses what any of them is worth.

## The six criteria and their weights

| # | Criterion | Weight | What it measures | How the points are given |
|---|---|---|---|---|
| 1 | Order-book liquidity | 30 | Depth, as the entries use the word: the money a BTC perpetual book has standing within 10 basis points of its mid, bids and asks counted together, taken as the median of the reading series rather than its high point. | Ten points is the ceiling and $10,000 of resting depth the floor, where the mark is nothing at all. Between the two the scale counts decades: every tenfold multiplication of the depth adds two points, which files $100,000 at two, $1 million at four and $10 million at six. |
| 2 | Market coverage | 20 | Two counts taken off a venue's own market list: the perpetual contracts on it, and how many of the asset classes standing among them are something other than crypto. | Six of the ten points ride on the length of the list, on a curve that pays the full six at 500 contracts and a shade under four at fifty. The remaining four are counted singly — a point per asset class beyond crypto, and the fourth one is the last that pays. |
| 3 | Trading cost | 18 | A round trip, meaning one fill into the BTC perpetual and one back out of it, both taken at the taker rate a new account is charged, the pair written in basis points. | A round trip of two basis points, or anything under it, has nothing deducted and stands at ten. Past two the deduction is half a point for every basis point, which empties the mark exactly at twenty-two; a dearer round trip is filed at zero all the same. |
| 4 | Security record | 14 | Entered from records a reader can open: the audit reports a venue has published, the page its bounty is posted on, and what the last twenty-four months of public reporting hold. | Two points an audit, and the count stops at the second, however many a venue lists; three points where a bounty stands open to anyone who finds something; three points where the twenty-four months behind an entry hold no loss of user funds |
| 5 | Custody and settlement | 12 | Read off a venue's own architecture, settlement and withdrawal pages: how much of a fill a chain is ever shown, and what the holder of the collateral can still do once the operator stops. | Four points where the chain is shown every trade, as a batch a rollup posts or as a proof it validates; two points where the chain is shown position snapshots at intervals and nothing in between; three points where the documentation sets out a withdrawal a user can force alone; two points where the contracts holding the collateral have been audited; one point where an account is a wallet signature with nothing kept behind it |
| 6 | Access | 6 | Taken from the terms a reader accepts and the sign-up page in front of them, then doubled, because the raw count only reaches five. | Two points where no identity document stands between a reader and a first order; one point where signing in is a wallet signature rather than an account; two points where the closed list goes no further than sanctioned states and one large market, one point where it reaches wider; the sum, five at most, is doubled, so this line is read on the same ten as the rest |


Nothing in that table is a wish about what a venue ought to be. Each line records what was counted
and what the count was worth, which is also what makes it arguable: a reader who would divide the
hundred points differently can take the six marks from the register below and reweigh them without
asking us for anything.

## Why the weights fall where they do

**1. Order-book liquidity, 30 points.** The question this reference asks is whether another venue can take the trade the biggest one is taking now. A venue whose book cannot absorb the order is not a substitute, whatever else it offers, so depth carries the heaviest weight and is the one criterion read off the book rather than off a page.

**2. Market coverage, 20 points.** The second reason a reader cannot move is that the market they trade is not listed. A venue with 52 markets is not a replacement for one with 324, and asset classes beyond crypto decide whether one account covers the whole book.

**3. Trading cost, 18 points.** Fees matter, but they separate the field less than they look. Across the nine venues scored here the entry-tier taker fee runs from nothing to five hundredths of a percent, and the tiers that charge nothing are paid for in latency instead. Material, and below liquidity and coverage.

**4. Security record, 14 points.** Moving away from the venue that holds the most open interest means accepting a shorter record, so audits, a bounty anyone can claim and losses inside the last two years are weighted above the settlement model rather than below it.

**5. Custody and settlement, 12 points.** Where a trade is recorded, and whether a withdrawal can be forced without the operator, is what this category claims to be about. It is documentation rather than measurement, so it sits under the two criteria a reader feels on the first trade.

**6. Access, 6 points.** Restrictions and account requirements filter who may use a venue rather than how good it is, so they carry the smallest weight and are given an entry of their own instead of being buried in a total.

## The terms, the ties and the last decimal

| Term | How this reference uses it |
|---|---|
| The scale | Six marks out of ten, weighted as above, set on a scale of 0 to 7 and cut to 1 decimal place. |
| Ties | Venues whose totals are equal after rounding share a place and are marked with an equals sign. |
| A loss of user funds | A loss of user funds is money taken in an exploit, lost in a venue-run vault through market manipulation, or lost through an outage of the venue, and not repaid in full. A liquidation caused by the market moving is not a loss. |
| The entry tier | What a new account pays by default on a venue's own interface, before any volume tier, rebate or discount is applied to it. |
| Resting depth | Orders standing in a public book within 10 basis points of the mid price, counted on both sides, read on a ten-minute clock across the two days to 20 September 2026. |

Rounding is the last thing that happens, and only the total is rounded: the six marks go into the
weighting as they were worked out. The marks printed in the register are cut to the same one
decimal, so a total rebuilt by hand from that row can land a hundredth away from the total beside
it. Where two venues still meet after rounding, neither is put above the other — they share the
place, and the entry sets an equals sign against both.

## Current scores

The register below is the whole scored field, in order of total, with the six marks in the order of
the table above. Documentation and terms were read on 18 September 2026, and the books over the
window set out above. Depth is the median of those readings, and the round trip is twice the
entry-tier taker fee on a $10,000 order.

| Venue | Score | Liquidity | Coverage | Cost | Security | Custody | Access | Median BTC depth | Readings | Round trip |
|---|---|---|---|---|---|---|---|---|---|---|
| Lighter | 5.8 | 6.5 | 9.2 | 10.0 | 7.0 | 10.0 | 8.0 | $16.9 million | 115 | $0.00 |
| Aster | 5.5 | 6.4 | 10.0 | 7.0 | 10.0 | 7.0 | 8.0 | $16.1 million | 117 | $8.00 |
| Extended | 5.5 | 5.7 | 9.6 | 8.5 | 10.0 | 7.0 | 8.0 | $7.4 million | 115 | $5.00 |
| edgeX | 5.3 | 6.9 | 8.0 | 6.5 | 7.0 | 10.0 | 8.0 | $29.5 million | 113 | $9.00 |
| Hyperliquid | 5.0 | 6.0 | 9.6 | 6.5 | 7.0 | 7.0 | 8.0 | $9.6 million | 118 | $9.00 |
| EVEDEX | 4.7 | 6.5 | 7.8 | 6.5 | 7.0 | 5.0 | 8.0 | $18.3 million | 115 | $9.00 |
| ApeX Omni | 4.7 | 4.5 | 7.7 | 6.0 | 7.0 | 10.0 | 10.0 | $1.8 million | 118 | $10.00 |
| Paradex | 4.4 | 1.0 | 7.0 | 10.0 | 10.0 | 7.0 | 8.0 | $32,811 | 118 | $0.00 |
| Aevo | 4.0 | 5.1 | 8.5 | 3.0 | 4.0 | 7.0 | 8.0 | $3.7 million | 117 | $16.00 |
| dYdX | 3.9 | 3.9 | 5.2 | 6.0 | 7.0 | 7.0 | 8.0 | $845,322 | 118 | $10.00 |


This is a register, not a verdict. It carries no places and settles no argument, because a total
only answers the narrow question the method asks and every entry asks a narrower one still. An
entry takes the venues its own question admits, orders them on the number that question turns on,
and leaves the rest of the field out. Where two rows print the same total, the register holds them
in the order of the unrounded figure — a hundredth of a point in one of the two cases here, four
ten-thousandths in the other — while an
entry, which does award places, gives the two of them the same one.

Readings is the number of times a venue's book answered. One answered 113
times against a median of 117 across the register, and the count is printed in the row rather
than averaged out of sight.

## Which venues the method admits

The condition, in the method's own words:

Venues that run a central limit order book, whether the book is matched off-chain, inside a rollup or inside a chain of the venue's own. A venue is scored only where its resting orders can be read from a public endpoint, because the heaviest criterion is measured rather than quoted, and only where its own documentation carries a BTC leverage cap of at least 40x, because the question an entry answers is whether an order can be moved to it at size. Nine venues clear both, and those nine fill the tables in the entries. Aevo caps BTC at 20x, so it is kept out of those tables; it is still read, measured and scored, and its row stands in the register on this page. Oracle-priced pools and vaults are described where a reader would look for them and are left out of the ranking.

That test takes in 10 of the 16 venues followed here and leaves out six, for
three different reasons. Four price from an oracle and fill against a pool or a vault instead of a
book, so there are no resting orders to read and the heaviest criterion would have to be guessed.
One has had the program behind it paused since April 2026, with its own address now pointing at a
fork running under another name. The last is infrastructure rather than a venue: one shared book
behind many front-ends, each of which sets the fee its own traders pay, so it has no single entry
tier to put in a cost column. All six are described in the entries where a reader would look for
them, with the reason printed beside them, and none is handed a total it did not earn on the same
terms as the rest.

## What sets the order of an entry

A score decides nothing about which venues an entry covers; the entry's own question does that.
What a score decides is the sequence, and only on the entries where no single measurement answers
the question better than the whole method does.

| Entry | Put in order by |
|---|---|
| Front page, GMX alternatives, Account requirements, Settlement compared | The total the six criteria come to |
| Hyperliquid competitors | The standing CertiK Skynet gave the venue, every venue read on one day |
| Perp DEX for large orders | The median dollars found resting within 10 basis points of the BTC mid |
| Leverage limits | The documented cap on BTC, read beside the notional it survives to |
| Cheapest perp DEX | What a $10,000 round trip is charged at the entry tier |


Where an entry is ordered on one measurement, the total usually stands beside it in the same
table, and the two disagree often enough to be worth watching: the venue with the deepest book on
[the large-orders entry](/perp-dex-for-large-orders) is not the venue at the head of the register,
and [the cost entry](/cheapest-perp-dex) puts two venues level at its top, only one of which leads
the register. Two entries print no total at all — [the leverage ladder](/perp-dex-leverage-limits)
and [the audit register](/hyperliquid-competitors) — because a first row carrying the lowest number
on the page contradicts itself; the totals for those venues are in the table above instead.

## What a total does not measure

A weight is a judgement, and there are six of them here. This method holds that a book able to
absorb the order matters more than where that trade is finally recorded. A reader who trades in
small size and cares most about being able to force a withdrawal out without the operator would
order those two the other way round, and that is a fair position — which is why every mark is
printed criterion by criterion and not only as a total.

Five of the six are read off documents: a venue's own for four of them, and published audit
reports, bounty pages and incident records for the fifth. Only liquidity is measured here. A
written term is not a tested one, and the method does not check that the written thing happens; it
records what a venue commits to in writing on the day the page was read, and scores that.

Outside the method altogether: how an interface behaves when the market moves, how long a
withdrawal actually takes that day, whether a support ticket is answered, and whether a reader will
be accepted at all. No account was opened and no order was placed, so none of it could be scored
honestly, and an invented criterion would be worse than a missing one. Perpetual futures are
leveraged contracts and the margin behind a position can go in full; a mark on this page is not
advice to put any of it anywhere.

## When the method changes

This is version 1.0, fixed on 22 September 2026. A weight, a rubric or the point
at which a curve stops paying can move only in a new version carrying a new date, and a new version
means the whole field is scored again from its inputs rather than nudged in place. A total that
moves because an input moved is an ordinary revision. A total that moves because the method moved
is a different document, and it says so.

Inputs move on their own schedule: a fee schedule is edited, a market list grows, an audit is
published, a book thickens or thins. That is why a figure is dated by the day it was taken and not
by the day an entry was last touched, and why the date under a heading moves only when something
under it did. Who compiles the entries, and how a correction reaches us, is set out in
[the front matter](/about).

Prepared by The Perp Reference maintainers for perpetual futures venues and the alternatives to the largest of them, last revised on 24 September 2026. Corrections and disputed figures: corrections@thekiddushclubnft.com.

The maintainers are paid to publish this reference, and a venue with an entry here may be the party that paid for it.