Adding a character to a random password buys you log₂(95) = 6.6 bits. Adding a word to a Diceware passphrase buys log₂(7776) = 12.9 bits — almost exactly twice as much, and a word is far easier to remember than a symbol. That is the whole argument, and it is why the advice everyone was given about symbols and digits was backwards.
How many equally likely possibilities a password was drawn from. It is a property of the process that generated it, not of the characters that came out, which is the one thing most strength meters get wrong.
bits = length × log₂(alphabet), or words × log₂(list) for a passphrase. Crack time is 2bits−1 ÷ guesses per second: half the keyspace, on average.
Ten tries a second at a login form against 10¹² a second offline on a fast unsalted hash: eleven orders of magnitude apart. A crack time with no attacker named is theatre.
Random words from a list of 7,776 (Diceware) at 12.9 bits a word. Easier to remember than characters, and only as good as the randomness that picked them.
A meter sees P@ssw0rd! and counts a capital, digits and a symbol. An attacker sees one dictionary word with the substitutions every cracking tool already applies. The characters look random; the process that produced them was not, and only the process counts.
log₂(7776) = 12.9 bits per Diceware word against log₂(95) = 6.6 per random character. One more word buys nearly twice what one more symbol does, and you can remember it. That is the whole case for passphrases.
Ten guesses a second at a login form and 10¹² a second against a leaked fast hash are eleven orders of magnitude apart. One number with no attacker attached is a marketing device.
Moving from unsalted SHA-256 to bcrypt takes an attacker from about 10¹² guesses a second to 10⁴: eight orders of magnitude, roughly 27 bits handed to every user at once. That is the site operator’s decision, not yours, and it dwarfs your password choice.
Forcing a symbol and a digit produced predictable results: the capital goes first, the digit and symbol go last. Attackers know the shape. NIST dropped composition rules and mandatory rotation from its guidance for exactly this reason.
A perfect 80-bit password used on two sites is only as strong as the worse-run of the two. Breach data is replayed against every other service within hours. Entropy protects against guessing; only a unique password protects against a breach elsewhere, which is what a password manager and two-factor authentication are for.
No. Everything on this page runs in your browser, there is no analytics on the site, and nothing is transmitted, stored or logged. You can check by opening developer tools and watching the network tab, or by disconnecting from the internet — the page keeps working, because it is all local.
Because most meters count character classes rather than guessability. They see a capital, a digit and a symbol and award points. P@ssw0rd! has all three and is among the first few thousand guesses any real attacker makes. Entropy comes from how the password was chosen, and a meter cannot see that.
Almost nothing. Replacing a with @ and o with 0 adds roughly one bit each against an attacker whose wordlist already applies those rules — which every real cracking tool does. You get the illusion of complexity and pay for it in memorability.
Guess rates only ever go up. The page assumes the fast offline rate keeps doubling every couple of years — a rough reading of GPU history, and a slider because it is a guess — and works out the year in which cracking your password would take under a day. One extra bit buys about two years; one extra Diceware word buys about twenty-five. It is the reason a password that is fine today is not fine in a decade, and the reason the argument for length never goes away.
Four is enough for anything rate-limited. Five is a sensible floor for an account that matters. Six puts you past what is feasible offline for the foreseeable future. The words must be chosen randomly — picking them yourself collapses the entropy, because people do not pick uniformly.
That the password was generated uniformly at random from the alphabet or wordlist you specified. Anything you type in is scored by the cheapest way to guess it: the string is cut into the patterns an attacker actually enumerates — dictionary words, leet spellings, keyboard runs, sequences, repeats, dates and plain digit runs — and each is priced by how many candidates it stands for. Because the built-in wordlist is only a few hundred entries, a second, pessimistic figure assumes every plausible word is in the attacker's dictionary. The lower of the two is shown, which is the one worth planning against. Not modelled: credential stuffing from a breach elsewhere, phishing, keyloggers, password reuse, or any attack that does not involve guessing. Guess rates are published 2026 benchmark figures for a single high-end GPU, multiplied by the rig you pick. They are the input most likely to be out of date, and they only ever move one way.
Releases in which this page changed, newest last. Derived from the archived copy of every release, not from notes written afterwards — so it reflects what actually shipped. Site-wide passes are left out; they are in the full changelog.