cryptonist
A bank vault standing open, its door thickness and locking mechanism visible

Wallets

12 or 24 words: what a longer seed phrase buys

A 12-word recovery phrase encodes 128 bits of randomness and a 24-word phrase encodes 256. Both are far out of reach of guessing, and the keys either one produces offer about 128 bits of security once a public key is visible, so the extra words do not make a wallet twice as safe. What they change is narrower: a longer phrase holds up better when part of it leaks and catches more copying errors, and it does nothing at all when the whole phrase is read, photographed or typed into the wrong page.

How the words are made

Both lengths come from the same standard, BIP-39, and the words are the last step rather than the first. A wallet starts by generating a block of random bits, which BIP-39 calls the initial entropy, 128, 160, 192, 224 or 256 bits long. It hashes that block with SHA-256 and appends the first few bits of the hash as a checksum, one bit for every 32 bits of entropy. Then it cuts the result into 11-bit pieces. Each piece is a number from 0 to 2,047, and each number picks one word from a fixed list of 2,048.

The arithmetic gives the familiar lengths:

Entropy (bits)Checksum (bits)TotalWords
128413212
160516515
192619818
224723121
256826424

Twelve and twenty-four are the two ends of the range. Wallets choose a default: MetaMask generates 12 words, Ledger’s documentation describes a 24-word phrase, and Trezor’s mentions backups of 12, 18 or 24 words.

The English word list is built for transcription. BIP-39 asks that the first four letters of each word identify it unambiguously and that near-twins such as “build” and “built” be left out; in the published list, no two words share their first four letters. The words carry no meaning of their own. They are a way of writing down a number, which is why BIP-39 describes itself as a means of carrying computer-generated randomness, and expressly not a way of turning a sentence someone made up into a wallet.

The checksum, and what it is for

Because 132 bits are spread across twelve words, the last word carries four bits of checksum alongside seven bits of entropy. Of all the possible sequences of twelve list words, 2,04812 or 2132 of them, only one in sixteen has a valid checksum. A 24-word phrase has an eight-bit checksum, so one sequence in 256 is valid.

The checksum is there to catch mistakes, and it does so imperfectly. If one word is copied wrongly, the corrupted phrase still passes a 12-word checksum about one time in sixteen and a 24-word checksum about one time in 256. BIP-39’s own list of shortcomings notes that the check is short and offers no help in correcting an error. It adds nothing secret either. It is computed from the entropy, so the number of valid 12-word phrases is exactly the number of possible 128-bit entropy values: 2128.

Why 12 words are already enough against guessing

2128 is 340,282,366,920,938,463,463,374,607,431,768,211,456. Testing a candidate is not free. A guess that passes the checksum has to go through PBKDF2 with 2,048 rounds of HMAC-SHA512, as BIP-39 specifies, then through the key derivation of BIP-32, and the resulting addresses have to be looked up on a chain.

There is a measured rate to scale from. The researchers who disclosed the Milk Sad flaw, described below, searched a space of 232 candidates, about 4.29 billion, and put that at a few days on an average gaming PC, at most. A 12-word phrase is 296 times larger, a factor of about 7.9 × 1028. Give an attacker a billion such machines, take “a few days” as three, and exhausting the space takes about 6.5 × 1017 years. NASA’s WMAP mission measured the age of the universe at 13.77 billion years. The search would run some 47 million times longer than that.

A 24-word phrase raises the space to 2256, larger again by a factor of 2128. That margin is real, but it protects the phrase rather than the keys. BIP-32’s security section assumes that working back from a public key to its private key means solving the elliptic-curve discrete logarithm problem, at about 2128 operations, and the standard notes elsewhere that its 256-bit keys offer about half their length in security. Every key in the wallet sits behind that same wall whatever the phrase length. Someone trying to break one exposed account would attack the curve at 2128, not a 24-word phrase at 2256.

Where the extra words do something

When part of the phrase is known

Why would length matter at all, if brute force against a whole phrase is hopeless at either length? Because a phrase does not always leak whole, and against part of one the length matters a great deal: the work depends on how many words are missing, not on how many there are.

Suppose someone has seen every word but not the order. Twelve words can be arranged in 479,001,600 ways, about a ninth of the 232 space the Milk Sad researchers searched in days. Twenty-four words can be arranged in about 6.2 × 1023 ways, roughly 279.

Suppose instead that half the words are known, in their places. Six unknown words out of twelve leave 2,0486 combinations, which is 266, and one in sixteen passes the checksum: 262 candidates, about a billion times the Milk Sad search. Twelve unknown words out of twenty-four leave 2132 combinations and 2124 valid ones, within a factor of sixteen of an entire 12-word phrase. A note holding half of a 12-word phrase gives away far more than a note holding half of a 24-word one. That is arithmetic about exposure, not storage advice; where to keep a phrase is its own question.

When copying it out

The longer checksum catches a single wrong word 255 times in 256, against 15 in 16 for twelve words. Twice as many words, on the other hand, means twice as many chances to make the mistake.

If quantum search becomes practical

The argument most often made for 256 bits concerns quantum computers. NIST’s report on post-quantum cryptography says Grover’s algorithm gives a quadratic speed-up for search, that nobody knows whether it will ever be practically relevant, and that if it is, doubling key sizes would be enough to preserve security. On that hypothetical, searching 128 bits would take on the order of 264 steps and 256 bits about 2128. The same report lists elliptic-curve signatures such as ECDSA as no longer secure against a large-scale quantum computer, through Shor’s algorithm. In that world the weak point would be the curve under every key, and phrase length does not touch it.

What 24 words do not protect against

The phrase being read. Nobody needs to search for a phrase they can see. A photograph in a synced gallery, a page that asks for the words, malware that reads them from a file: each hands over the whole phrase, and a whole phrase of any length restores everything. Canada’s Anti-Fraud Centre says as much in a bulletin on wallet fraud: anyone who obtains the phrase may be able to move the funds without the victim’s phone, computer or device. How phrases are actually taken is the subject of can a seed phrase be hacked.

A generator that never put the randomness in. The word count shows how many bits the format can carry, not how many the software supplied. The two Milk Sad victims whose losses started the investigation had used 24-word phrases created on air-gapped laptops. The tool that made them, the bx seed command in Libbitcoin Explorer 3.0.0 to 3.6.0, used a generator that, in the words of the NVD entry for CVE-2023-39910, restricts the internal entropy to 32 bits regardless of settings. Twenty-four words drawn from 32 bits of randomness are 32 bits of security written out at length. In July 2026 Coinkite disclosed that seeds generated on its Mk4, Q and Mk5 Coldcard devices before fixed firmware had about 72 bits of entropy where 128 were expected, and that seeds from the older Mk2 and Mk3 were affected more severely.

Losing it. Length has no bearing on whether a phrase survives a fire, a house move or the death of the only person who knew where it was kept. Losing a seed phrase covers what can and cannot be recovered.

A forgotten passphrase, or a bad signature. An optional passphrase changes which wallet the phrase opens, and is covered in the 25th word. Neither the phrase nor its length has anything to do with what you approve when you sign.

Choosing, and changing your mind

The length is fixed when the phrase is generated, because the words encode that particular entropy. Adding words to an existing phrase does not lengthen it; it produces a different sequence, almost always one whose checksum fails. Moving from 12 words to 24 means generating a new phrase, which means new keys, new addresses and a transfer of the funds.

Either length controls the same thing: every account the wallet derives from it, on every chain. How a seed phrase differs from a private key sets out that tree. Your first crypto wallet covers the habit that matters more than the word count, which is restoring the phrase into a second wallet before trusting it with anything.