as text tokens
—
chars ÷ 4
Glyph
A language for agents.
Glyph encodes your context into the cheapest form your model can actually read — a dense image set in a learned script where that wins, plain text where it doesn’t — and shows you the per-reader token math either way.
Specimen — a real render, not a mockup
What would your context cost?
as glyph · tiled billing
—
Gemini / Qwen class
as glyph · pixel billing
parity or worse
Claude / GPT class → Glyph returns text
The grammar picks per reader: only a strict win (ratio > 1.0) gets the image. Everything else gets text — you can’t be overcharged by a false win.
The honest ledger
| Reader class | Billing | Glyph vs text | What you get |
|---|---|---|---|
| Gemini / Qwen class | flat tile | 1.06–2.33× fewer tokens | glyph image — wins only on dense sets, and only when strictly > 1.0× |
| Claude / GPT class | pixel area | parity or worse | plain text — always |
| Unknown reader | assumed pixel | — | plain text — ambiguity never manufactures a win |
| Verbatim content (code, hashes, quotes) | any | excluded | plain text — vision models paraphrase; Glyph is for comprehension, not citation |
These numbers come from an adversarial falsification harness, not a benchmark we liked (docs/substrate_falsify.md). The 15× figure you may have seen is legibility per page versus a normal-font screenshot — an image-to-image comparison, never versus text tokens. The economic wedge is uncached, first-send context: RAG results, one-shot agent context, batch document QA.
Three layers, one language
01 · script
Learned letterforms
A 20KB bitmap font trained against one vision model that transfers to others — trained on Qwen-3B, read by Qwen-7B (5/5) and Claude (4.7/5 MCQ). Denser than any system font a vanilla renderer can use.
02 · grammar
Reader-aware encoding
Every response is costed against your reader’s actual billing — tile, pixel, or weak — and the cheapest legible substrate wins. The math ships with the payload.
03 · corpus
Fresh content, already spoken
A time-series index of the internet since your model’s cutoff — 460,000+ documents across HN, arXiv, GitHub releases, CVEs, papers, frontier labs, markets — queryable in Glyph from day one.
Speak it from anything
# encode your own context (the language) POST https://glyphapi.dev/v1/encode { "text": "…", "reader": "gemini-2.5-flash" } → cheapest substrate + per-reader token manifest # search the fresh corpus, answered in Glyph POST https://glyphapi.dev/v1/retrieve { "query": "prediction market odds", "substrate": "auto" } # MCP — Claude Code, Cursor, Zed, anything https://glyphapi.dev/mcp/ (tools: encode, retrieve_auto, glyph_search, …)
Free taste without an account: POST /v1/probe (one real query, 1/IP/day) · GET /v1/glyph/sample.png · then a 20-call trial key, no card, no wallet.