The Model in One Sentence
The meter estimates how many passwords an attacker would have to try before reaching yours, then divides by a guessing speed of ten billion per second. Everything else on this page explains those two numbers.
The size of the search is measured in bits. Each extra bit doubles the number of possibilities: a 50-bit search is about a quadrillion candidates, and a 60-bit search is a thousand times larger. The details view shows this number as “~60-bit search estimate”.
The Assumptions
- The attacker already has the hash. A service stores a scrambled form of your password, called a hash. The model assumes that copy was stolen in a breach, so the attacker can test guesses privately. Nothing slows them down: no lockout after five tries, no rate limit, no alert.
- The hash is a fast one. Ten billion guesses per second is roughly what one or two high-end graphics cards manage against a general-purpose hash such as SHA-256. A single NVIDIA RTX 4090 has been benchmarked at about 22 billion SHA-256 guesses per second, and faster still against older hashes such as MD5 (about 164 billion).
- The attacker has average luck. They might hit your password on the first try or the last. On average it takes half the search, so the estimate uses half.
- The attacker tries the obvious things first. Common passwords, keyboard runs, dates, repeated characters, and your own username or email are counted as a handful of guesses each, not as random characters. That is why
Password2024!rates as very weak despite its length and symbols. - Hardware stays as it is today. The estimate does not project faster machines. Every doubling of attacker speed removes one bit, which halves every time shown.
Demanding on purpose
These assumptions favor the attacker. A password you save for another website may be stored there with a fast hash or a slow one, and you cannot tell which. Assuming the fast case means a “Strong” rating holds up either way.
What Would Break It in a Day?
Turn the question around: for a given password strength, how much guessing power would finish the job in 24 hours? The hardware column uses the benchmarked SHA-256 speed of one RTX 4090 as its unit.
| Search size | Randomly generated example | Time shown in OWL | To break it in a day |
|---|---|---|---|
| 40 bits | 8 lowercase letters and digits | 55 seconds | About 6 million guesses a second. An ordinary laptop. |
| 50 bits | 10 lowercase letters and digits | 16 hours | About 7 billion a second. One high-end graphics card. |
| 60 bits | 10 letters and digits, mixed case | 2 years | About 7 trillion a second. Around 300 graphics cards: a rented cluster or a dedicated cracking rig. |
| 70 bits | 12 letters and digits, mixed case | About 2 thousand years | Around 300,000 graphics cards. The scale of a large data-center fleet. |
| 80 bits | 6 words, or 13–14 mixed-case letters and digits | About 2 million years | Around 300 million graphics cards. More than any one organization is known to operate. |
| 100 bits | 16 characters including symbols | Past the Sun’s red-giant phase | Hundreds of trillions of graphics cards. Not buildable. |
So under the model, the answer to “what could break this in a day?” is: one graphics card for anything near 50 bits, a serious cluster near 60, and a nation-scale effort near 70. From about 80 bits upward, no known collection of general-purpose hardware gets there.
For a sense of the ceiling, the Bitcoin network is the largest hashing effort ever assembled, on the order of 1021 SHA-256 computations per second. Its machines are single-purpose chips that cannot be pointed at passwords. If they could, an 80-bit password would fall in minutes, a 100-bit one would still take about twenty years, and a 128-bit one billions of years. That is why generated passwords aim well above 80 bits.
How the Ratings Follow
The word beside the time is set by the time, so the two cannot disagree.
| Rating | Search size | Estimated time |
|---|---|---|
| Very weak | Under 47 bits | Under an hour |
| Weak | 47 to 59 bits | An hour to about a year |
| Fair | 60 to 69 bits | Years to centuries |
| Strong | 70 to 79 bits | Thousands of years or more |
| Very strong | 80 bits and up | Millions of years or more |
A password built mostly from an obvious pattern is held to a lower rating even when its length suggests more. Beyond a few billion years the meter stops counting and names a cosmic milestone instead, because a number with thirty digits says less than “after the longest-lived stars burn out”.
Why Real Sign-Ins Are Usually Slower
The model describes the worst reasonable case. Two things usually make an attacker’s life much harder.
- Online guessing is slow. Without a stolen hash, an attacker has to try passwords against the live sign-in page, where rate limits allow a few guesses, not billions. At that pace even a mediocre password that is not on a common-password list lasts a long time.
- Good services use slow hashes. Password-hashing functions such as bcrypt, scrypt and Argon2id are built to be expensive. The same graphics card that tries 22 billion SHA-256 guesses per second manages about 184,000 against bcrypt at its lowest benchmark setting, and about 7,000 against scrypt. At 7,000 guesses per second, the 50-bit password that fell in 16 hours would take that one card over two thousand years.
Your OWL account is in the slow category. Sign-in uses OPAQUE, so your password never leaves your device, and each guess must pass through Argon2id using 64 MiB of memory. An attacker could not begin guessing offline without first stealing both the account records and the server’s own secrets. The meter still shows the fast-hash estimate, because the same meter rates passwords you store in OWL Pass for other services, and because a password that survives the harsh model survives the gentle one.
What the Estimate Cannot See
- Reuse and leaks. A password already exposed in another site’s breach is tried first, whatever its length. The estimate assumes the password is unique to one place.
- How you chose it. The meter sees only the characters, so it discounts anything typed by hand: people rarely choose at random, and a typed password scores lower than a generated one of the same shape. It still cannot tell a truly random string from a memorable one that happens to look random, such as a song lyric’s initials. Only a generator knows the real figure, which is why OWL’s password generator reports its exact bits instead of an estimate.
- Attacks that skip guessing. Phishing pages, malware that records keystrokes, and someone reading a sticky note all bypass strength entirely. Two-factor sign-in and a password manager address those.
- Future machines. A large quantum computer running Grover’s algorithm could in theory search as if a password had half as many bits. No such machine exists, and it is another reason generated passwords carry a wide margin.
The practical takeaway
Length from a generator beats cleverness. Six randomly chosen words, or sixteen random characters, put a password in the “not buildable” rows of the table above. Use a different one for every account and let a password manager remember them.
References & Further Reading
- Chick3nman, Hashcat v6.2.6 benchmark on the NVIDIA RTX 4090 (2022) — source of the per-card guessing speeds quoted here.
- NIST, SP 800-63B: Authentication and Authenticator Management — current guidance on password length, blocklists and rate limiting.
- IRTF, RFC 9106: Argon2 and RFC 9807: OPAQUE — the memory-hard function and sign-in protocol OWL uses.