nc://model-tiers

The Claude model hierarchy, and the leaked prompt archive behind it, quoted with line numbers.

The fact the archive states and no vendor page does: above Opus sits a Mythos tier, and Claude Fable 5.2 is a Mythos-tier model with extra safety measures for biology and cybersecurity. It shares one underlying model with Mythos 5.2. The generation before says the same of 5.1, in its own file - which is the point: one document can be a fake, two that agree a generation apart are harder to fake.

Get it

Tell the agent "add the model tiers cell", or run nodecode add nc://model-tiers.

Use it

Ask in your own words: what is above Opus, is Fable the same model as Mythos, what did the export controls do, what does Cursor's prompt say, how does Codex describe itself. The model reads the same archive you do.

• model_tiers_files - the archive: thirteen documents, what each one is, its vendor, size and digest. Filter by vendor (Anthropic, OpenAI, Z.ai, Meta, Cursor, Cline) or a word.
• model_tiers_list - the Claude ladder, tier by tier (mythos, opus, sonnet, haiku).
• model_tiers_get - one model whole: where it sits, whether a person can have it, its API string, and the archive's sentences with document and line.
• model_tiers_search - any word, over the models, the tiers and the documents' own summaries.
• model_tiers_quote - one document's own lines: every line carrying a word, numbered, or a range as 116-122. This is the one verb that fetches, once per document.

In chat: /model-tiers for the ladder and the timeline, /model-tiers files for the archive, /model-tiers fable for one model, /model-tiers source for the archive's head and a digest, /model-tiers lines 116-122 for the prompt's own words.

From Lisp, with the cell on:

(nodecode-model-tiers:models-matching :tier "mythos")
(nodecode-model-tiers:files-matching :vendor "OpenAI")
(nodecode-model-tiers::quote-text :file "cursor-2" :lines-text "1-12")

What it does not do

It does not ask a vendor anything, and it does not crawl. The hierarchy and the index are answered from data.lisp with no request, no key and no account. Only model_tiers_quote reaches the network: it fetches one document from the archive once, caches it under ~/.cache/nodecode-model-tiers/, and reads the cache after that.

Read it with care

A snapshot: elder-plinius/CL4R1T4S at 93a58476, 9 November 2026, read 11 November. Every document carries its size, its line count and its sha256, and every quote carries a document key and a line number, so a person can check any sentence against the file it came from - that is what model_tiers_quote is for.

Where the archive and a vendor page disagree, this cell gives the archive's reading of itself, and says so in every answer.

None of it was announced. It was read out of files.

Set it up

Nothing is required.

Credit

The documents are elder-plinius's collection, in CL4R1T4S; the words in them are each vendor's. The code is MIT, ours. See NOTICE.