Knowledge baseSourcesMetehan Yesilyurt

Reverse-Engineering the OpenAI's GPT-5 Tokenizer: What 200,000 Tokens Reveal About AEO/GEO

Author Metehan Yesilyurt Date 2026-02-13 Model state o200k vocabulary (GPT-5 tokenizer), static analysis - independent of model behaviour Open source →

Key findings

Yesilyurt dissects the 200,000-entry vocabulary of the GPT-5 tokenizer: some brands exist as a single token, others split; URLs always split; prose is the cheapest format in tokens, tables the most expensive; over a thousand token slots are reserved for tools. His thesis: what is cheap to tokenize is easier for the model to process and reproduce.

Why this source matters for the project

As a mechanics hypothesis behind two of our own findings: first, ChatGPT partly phrases fan-out queries in English even for German questions (Share of English fan-outs) - English text tokenizes more efficiently in the o200k vocabulary, which would be a motive. Second, the single-token thesis makes it plausible why well-known brands surface so effortlessly in self-written queries. Both remain conjecture, but as the only mechanics explanation of this kind it belongs in the register.

Checked against our data

None of this is observable in the client capture - the tokenizer works before everything we see. The claims are therefore all unverifiable, but externally checkable (tiktoken) and not in contradiction with our data.

ClaimStatusEvidence
Brands like Google, Amazon, Reddit are single tokens; OpenAI and ChatGPT split into severalunverifiableA tokenizer property, independently checkable with tiktoken - but not observable in our setup (HAR captures). No contradiction with our data.
URLs are never atomic tokens, they always split into fragmentsunverifiableSame situation: static vocabulary property, outside our measurement window.
Format efficiency: prose ~5.9 characters per token, markdown 4.8, JSON 4.0, tables 2.7unverifiableNot checkable client-side; our snippets (202 characters index, 131-158 scrape) say nothing about token costs of the source page.
Around 1,075 reserved tool slots in the vocabulary point to planned tool expansionsunverifiableA vocabulary observation; our captures only show the tools actually active.

Related

yesilyurt-rrf wells-2026-02-12