Show HN: Unigram: encode bytes as words that cost one LLM token

Unigram, a new Rust crate released on crates.io, encodes bytes as words that each cost exactly one LLM token, making an N-byte value cost exactly N tokens under Claude and other models. The bijective codec, available via `cargo add unigram`, offers canonical parsing and tolerant recovery, with benchmarks showing it beats hex, base64url, and base58 at all sizes under Claude, and wins at 4 bytes under GPT-4o's o200k vocabulary.

A bijective codec between bytes and words that cost exactly one LLM token. cargo add unigram · crates.io https://crates.io/crates/unigram · docs.rs https://docs.rs/unigram · CHANGELOG /bleugreen/unigram/blob/main/CHANGELOG.md php a14ed61a - password email share building 8623a771b764ce50bb85371ff65aebe9 - links change points high random found season events region light const case users field table support An identifier becomes something you can read. Say it out loud, carry it across a room or between two windows, tell it apart from its neighbour at a glance, recognise it again an hour later — the ordinary things a name affords. Ids spend their lives in prompts, logs, and error messages, being looked at; this makes that free. One word is one byte and one token, so a value costs exactly as many tokens as it carries bytes — flat, for every value, with the spaces between words costing nothing. The four words above carry 32 bits in 4 tokens; the sixteen carry 128 in 16. js use unigram::{UnigramId, CheckedUnigramId}; let id: UnigramId<4 = UnigramId::try random ?; // 32 fresh bits, 4 tokens println "{id}" ; // "password email share building" let returned = UnigramId::<4 ::parse &text ?; // canonical: exact let salvaged = UnigramId::<4 ::recover &text ?; // tolerant: forgives a round trip // One extra word of CRC-8, when a mutated value must not pass as a valid one. let checked: CheckedUnigramId<4 = CheckedUnigramId::try random ?; The bytes are the value; the words are how it is displayed and parsed. Holding it that way means the length is part of the type, equality is byte equality, and there is no question of what format a given value is in — the question a string-shaped API cannot answer and has to guess at. Free functions encode , decode , decode recovered , try mint are there for variable-length payloads. parse is canonical: lowercase alphabet words, single spaces, nothing else. One accepted spelling per value, which is what belongs where a value is about to be trusted. recover forgives what a round trip through a model does — case, separators, line wrapping. It reads the whole input, so isolate the candidate first. Both refuse an unknown word and name it. One word is one byte and one token, so an N-byte value costs exactly N tokens, the same for every value. Mean tokens under Claude, with the worst of 200 deterministic payloads in parentheses: | encoding | 4 bytes | 8 bytes | 16 bytes | 32 bytes | |---|---|---|---|---| unigram | 4.0 4 | 8.0 8 | 16 16 | 32 32 | | hex | 6.0 9 | 11.3 15 | 21.7 27 | 42.6 52 | | base64url | 6.3 9 | 10.8 14 | 21.3 25 | 41.2 48 | | base58 | 6.6 9 | 10.9 13 | 21.2 26 | 42.0 47 | The parenthesised figure matters as much as the mean. Every other encoding's cost swings with the value, so a budget built on one has to assume its worst case; this one is known before the value is minted. Hex loses everywhere, at every size, in every family. The GPT vocabularies have memorised base64 fragments, which changes that ranking above 4 bytes — under o200k , base64url averages 29.5 tokens for 32 bytes against a flat 32, while unigram still wins at 4 bytes 4.0 against 4.5 . Nonce and correlation-id widths are what this was built for; a 32-byte digest is a worse fit, at 224 characters and no token margin left under GPT. One token per byte holds space-prefixed and bare , so a value costs exactly N at the start of a string, after a space, in JSON, and mid-sentence. The only surcharge is punctuation immediately before it. Measured for a 4-byte value against an ideal of 4, sweeping all 256 entries through the opening and closing positions, worst kept: | context | GPT-4o | GPT-3.5/4 | GPT-3 | GPT-2 | Llama | Claude | |---|---|---|---|---|---|---| | start of string | +0 | +0 | +0 | +0 | +0 | +0 | in prose, X. | +0 | +0 | +0 | +0 | +0 | +0 | JSON "id":"X" | −1 | +0 | +0 | +0 | +0 | +1 | | after a newline | +0 | +0 | +0 | +0 | +0 | +1 | after id: | −1 | −1 | −1 | −1 | −1 | +0 | markdown X | +1 | +1 | +1 | +1 | +1 | +0 | after | +1 | +1 | +1 | +1 | +1 | +0 | So: one token per byte, plus at most one for punctuation immediately before it — a constant, never scaling with the payload, and negative where the context ends in a space the value absorbs. That is a property of the table, and it was not free. 0.2.0 shipped 22 entries costing two or three tokens bare, so a value opening with council cost N+2 at the start of a string — and its verifier tested one payload whose opening word happened to be cheap. Both are fixed. The sweep is why the claim needs no exception list. BIP39, Diceware, the PGP word list, and what3words all predate this and all map data to words. None was chosen for tokenizers, and it shows. BIP39 is the closest comparison — 2048 words, which would be 11 bits each if they were all single tokens: | wordlist | words | single-token both ways, all families | usable alphabet | |---|---|---|---| | BIP39 | 2048 | 349 | 256 → 8 bits/token | unigram | 256 | 256 | 256 → 8 bits/token | Only 349 of BIP39's 2048 survive the filter, and Claude is the binding constraint at 366. Round 349 down to a power of two and a BIP39-derived encoding lands on exactly 256 entries and exactly 8 bits per token — the same density, from a list that also has no bare-cost or surrounding-context guarantee. BIP39 optimises for a different thing, and does it well: unique four-character prefixes and human-transcription distance, for seed phrases read off paper. That is worth having. It is not what makes a word cost one token. Tokenizer vocabularies hold their canonical word entries space-prefixed, so the space between two words is absorbed into the word that follows it and costs nothing. No other separator is free. Measured across all five families, an eight-byte value: | separator | GPT-4o | GPT-3.5/4 | GPT-3 | GPT-2 | Llama | Claude | |---|---|---|---|---|---|---| | space | 8 | 8 | 8 | 8 | 8 | 8 | . | 8 | 8 | 15 | 15 | 15 | 15 | - | 11 | 9 | 15 | 15 | 15 | 15 | , \n | 13–15 | 12–15 | 15 | 15 | 15 | 15 | The join would cost almost as much as the payload. Encoded values travel inside quoted strings in practice, where embedded spaces are free — and recover accepts every one of those separators anyway, so a value that comes back joined differently is not lost. 256 entries of lowercase ASCII English, 4 to 10 characters, under five constraints: One token, space-prefixed and bare, under every tokenizer the verifier pins: OpenAI's r50k base , p50k base , cl100k base , o200k base ; the hf-internal-testing/llama-tokenizer SentencePiece artifact at revision d02ad6cb ; and ctok 1.0.0's "5.0" counter, an offline reconstruction of Claude's tokenizer rather than Anthropic's own — checked against Anthropic's official count tokens endpoint on claude-opus-5 for all 256 entries, spaced and bare, where it agrees exactly verify-claude.py reruns it . Those exact artifacts are the claim, not every model that shares a name, and in particular not Llama 3, which tokenizes with tiktoken rather than the SentencePiece model checked here. No two entries within one character edit, and none a prefix or suffix-derivative of another. A slipped character, a dropped suffix, or a completed word lands outside the alphabet rather than on a different valid entry. Nothing charged — no death, violence, race, gender, religion, or politics. These strings surface unbidden in transcripts, logs, and user-facing errors. No function words. A value made of that , which , and would reads as damaged prose rather than as a name. Frozen. Byte n is ALPHABET n , all 256 slots are occupied, and changing an entry changes what every previously issued value decodes to. A test pins the table's digest. Nothing in an encoded value says which table produced it, so a system that stores these must record FORMAT VERSION alongside them. The crate depends on nothing but the OS CSPRNG, at runtime or under test, and never tokenizes. cargo test covers the codec and the table's structure — sorted, unique, lengths, edit distance, prefix and suffix relationships, the frozen digest, and an exhaustive sweep of every single-word substitution against the check word. It says nothing about cost. Every number on this page is printed by verify-alphabet.py , which reads the alphabet straight out of src/lib.rs , re-measures every entry against all five families both space-prefixed and bare, and sweeps all 256 entries through the opening and closing positions of every context — with dependencies and the tokenizer revision pinned exactly: uv run verify-alphabet.py Run it after any edit to the table. A green test suite alone establishes none of what this crate is named for. verify-claude.py is the audit for the one measurement that is a reconstruction rather than a vocabulary: it re-checks the Claude column against Anthropic's official count tokens endpoint and reports any entry where the two disagree. Needs ANTHROPIC API KEY ; roughly 300 calls with --bare . It last ran clean on every entry. MIT.