Token Space Fonts builds a font where every LLM token is the same width, so a developer can see how the model splits text

You cannot see the units your model is billed in, and that is the whole friction the tool below removes. Ampdot has published Token Space Fonts, a browser page that compiles a single TTF font in which every token produced by a chosen tokenizer occupies the same width on screen. Words that share a token render the same width. A word split into two tokens takes twice the space. The tokenization is visible in the text itself, rather than in a separate counter.

The tool went up on 25 September 2026 and reached 58 Hacker News points by the following day. The compiler is written by ampdot and runs entirely in the browser: no upload, no server. It accepts a base font (Inter, Roboto, JetBrains Mono, SF Compact, Bebas Neue, Caveat, or an uploaded TTF, OTF, WOFF or WOFF2) and a tokenizer preset, then writes a TTF with the tokenizer's shaping rules built in. The output installs like any font.

The tokenizers it covers

The presets are DeepSeek V4.1 Flash, ctok versions 4.8, 4.7 and 3, OpenAI's o200k_base, cl100k_base, p50k_base and r50k_base, Kimi K3, GLM-5.3 experimental, Llama 3 experimental, Trinity Large Thinking, Laguna M.1, Qwen 3.6, Gemma 4 and Gemini 3.5 Flash. A Hugging Face repository link or a tokenizer.json upload is also accepted, though ampdot says arbitrary uploads have not been through the preset-specific audits.

Claude is covered through ctok, an unofficial approximation. Ampdot names this itself in the compatibility notes: "ctok is an unofficial approximation, not a verified implementation of Claude's tokenizer". So the Claude view is close, not exact.

What ampdot measured

Ampdot reports a directed whitespace audit of 17,904 BPE checks across six font binaries and 2,355 ctok cases against local tokenizer references, plus 60 browser widths and 20 wrapping samples. Trinity Large Thinking and Laguna M.1 passed all 2,984 directed whitespace checks. GLM-5.3 and Llama 3 reached a 64-byte, 32-round shaping limit on 353 and 354 of 2,984 checks respectively. DeepSeek had 59 ID mismatches out of 2,984, all in the U+001C to U+001F control-separator range. The page lists each known gap by name rather than describing the tool as complete.

Ampdot also names the shape of the compromise: browser shaping runs, line breaks, normalization and complex scripts can move the result, and the output is a token font rather than an exact whole-message token counter.

Where it plugs in

The page ships two integrations alongside the compiler: a Vesktop theme for Discord that applies the compiled font to a chosen user's messages, and a userscript for web Slack that applies it to messages from specified display names or member IDs. Both stay local; the userscript makes no network requests and does not change the message text.

The reason to install one

A developer running an agent, tuning a prompt or costing a batch of calls works in tokens whether they see them or not. Reading token boundaries in a normal editor takes a separate tool and an extra step. A font that displays them in place puts the same information in front of your eyes while you write, and does it in whichever chat, editor or terminal will render the font. That is a small change in workflow and a real one: the units a model prices, splits and hallucinates on stop being invisible.

Source

Primary source: Token Space Fonts by ampdot.

Discussion: Hacker News thread.