GPT-6.2 Astra and Fable 5.5: What the Records Show

Three model names ran through the internet over 36 hours: gpt-6.2-astra, Claude Mythos 5.5 and Claude Fable 5.5. None of them is a product you can buy today, and that is not a judgement — it is what the four public records that decide the question said when we fetched them on 2 November 2026: OpenAI’s model docs and API changelog, Anthropic’s live Mythos and Fable pages, Azure’s model catalogue, and OpenRouter’s model API. Two of the three names were never anything at all. The third was real, briefly: six version strings in one component of one page, inside an HTML document whose logo file never changed.

Disclosure: this site runs on Cloudflare Workers and D1 (wrangler.json), has no affiliate relationship with OpenAI, Anthropic, Microsoft, LiteLLM, OpenRouter or Hugging Face, and nobody sent us a screenshot. Every count, string and byte size below comes from a request we made ourselves on 2 November 2026; the URLs at the end let you re-run each one.

What the records say today

That table is the whole story on availability. Everything else is about how three names got out of it.

The Fable 5.5 specs are a deleted row copied from a live one

The “Fable 5.5 is on Azure” claim has a single, datable origin, and it is not Azure. On 13 October 2026 at 12:27:21 UTC a bot named berriai-litellm-provider-info-sync committed feat(azure): add azure_ai/claude-fable-5-5 (#46201) to LiteLLM’s price list — 33 added lines in each of two files. On 14 October at 16:16:59 UTC, 27 hours 49 minutes later, devin-ai-integration[bot] committed fix(cost-map): remove azure_ai/claude-fable-5-5, which Azure does not list (#46428). We read both patches. The deleted record said:

Not one number was new. The “leaked specs” are Fable 5.1’s shipping record with a different key pasted on front — down to a retirement date for a model that did not exist. Two fields were extra: a duplicated max_tokens, and a source naming the Azure management API’s model list for one subscription in one region — …/locations/eastus2/models, API version 2024-10-01, with the Microsoft.CognitiveServices namespace elided. Azure’s own published catalogue names exactly two Fable SKUs — claude-fable-5 and claude-fable-5-1 — and its region table also puts both in eastus2 only.

The tell runs one step further back. claude-fable-5’s own Azure row prices cache reads at 1e-06 — $1.00 per million — while 5.1’s is $0.25, and Anthropic’s Fable page says so in words: cache reads are “$0.25 per million tokens, 75% less than Fable 5”. The deleted row carried $0.25, so it was copied from the 5.1 row, not derived from anything. LiteLLM’s own 1 September commit that added 5.1 says the same trick happened one generation earlier: “Specs match Fable 5 (1M context, 128K output, $10/$50 per MTok), except cache reads land at $0.25 per MTok.”

Why this matters more than a retracted rumour: LiteLLM’s file is a cost map. A gateway that has an entry for a model also has prices for it, and teams bill against those numbers. The failure mode here is not that someone predicted a model early. It is that an automated sync wrote a plausible row for a SKU its own source did not support, and it took a second bot 28 hours to notice.

Mythos 5.5 was in the HTML — and the archive dates it

The Anthropic half was reported as “a screenshot of the page code”. We could not find an archived copy in the reporting, and one write-up stated plainly that nobody had produced one. Wrong: the Internet Archive has ten crawls of anthropic.com/claude/mythos between 20 October and 3 November, and we counted the version strings in each.

So the string was served on 31 October and gone by 1 November — a window of at most 6 hours 16 minutes, and that bound is only as tight as the archive’s crawl rate. All six occurrences are in one place: a hidden h1 and the two wordmark images’ alt text, repeated once more inside the Next.js payload. Verbatim, from the 31 October crawl:

<h1 class="kMpfDa_hidden-header">Claude Claude Mythos 5.5</h1> <img alt="Claude Claude Mythos 5.5" loading="lazy" width="1040" height="141" …>

Three things this settles. The screenshot was genuine — the text was there, six times, before the first public post about it. The count arithmetic is exact: 37 + 6 = 43, i.e. the six places that read 5.5 on 31 October are six of the 43 that read 5.1 now, and nothing else on the page changed count. And the wordmark <img> src is byte-for-byte the same two SVG files that serve 5.1 today (996bdedc…-1040x141.svg, 1e46f1ab…-184x156.svg) — a version string was swapped in a component; no logo, no price, no spec, no benchmark on that page moved.

Read the live Fable page for the same lesson in its permanent form. Its masthead currently says Claude Claude Fable 5 while every sentence below it describes 5.1. Front-end strings on marketing pages drift, in both directions, and they drift without engineering meaning anything by them.

GPT-6.2 Astra: one image set, and a changelog that stops on 30 October

The OpenAI half rests on screenshots of the Codex usage page. TestingCatalog, which has looked at more of them than anyone, reports the ID appearing only on the 31 October bar, at wildly different shares by account — 1.6% of one week, 19.8% of another’s, 37.7% of one day — with the model picker still reading GPT-6 Astra. That is consistent with a staged test. It is also one image set: the same screenshots recirculated through several accounts, and the API-response fragment in circulation carries no endpoint, request ID, header or timestamp, so nothing in it ties a token to a build.

Against that, the record is unambiguous and checkable in sixty seconds. OpenAI’s changelog’s newest entry is 30 October (GPT-6 Rosalind, life-sciences reasoning for approved organisations); 26 October shipped GPT-6.1 Terra. The models docs price four flagships, and 6.2 is not among them. OpenRouter sells gpt-6-astra and gpt-6-astra-pro, not 6.2.

Two things in the rewrites are contradicted by those pages. Claims that OpenAI “skipped GPT-6.1” are wrong on OpenAI’s own catalogue: gpt-6.1-sol ($2 / $10) and gpt-6.1-terra ($1 / $5) are shipping models listed under “flagship”. What OpenAI cancelled was GPT-6.1 Astra, in late September, on alignment grounds — we covered that here — and a 6.2 Astra would be the retrained successor, which is a different claim, and one OpenAI has not made.

What the copies added that no record supports

The version that travelled furthest added capabilities nobody has evidenced: Mythos 5.5 described as “safeguard-free”, able to root a Linux kernel in 53 minutes; Fable 5.5 said to be “already with partners”; a “leaked weights” repository with thousands of downloads. On the last one we ran the Hugging Face API directly, because a fake weights repo is the part that can actually hurt a reader. Searching mythos returns 100 repositories, and the recent ones are not Anthropic’s: fableforge-ai/Mythos-V2-9B-GGUF (created 19 October, 8,070 downloads) is tagged qwen3.5, licensed apache-2.0, with base_model: fableforge-ai/Mythos-V2-9B and 25 GGUF quants. diablomike20/Claude-Fable-5.5 (created 28 October, 136 downloads) is tagged library_name: wan2.2 and ships an 8,988,111,328-byte file named claude-fable7u-se.gguf plus a Modelfile and two benchmark PNGs. Neither lab has published weights for anything named 5.5 or 6.2 that we could find.

Why the window still matters

Release cadence inside these families has run about two months: Fable and Mythos 5.1 on 1 September; GPT-6 Astra on 3 September; GPT-6.1 Terra on 26 October; GPT-6 Rosalind on 30 October. The leaderboard pressure is real and public — Artificial Analysis’s index today puts Gemini 4 Carbon (high) at 59, Claude Opus 5.5 (max with fallback) at 58, GPT-6 Astra (max) at 57. And Anthropic has an obvious external clock: its S-1 announcement page is live, listing on Nasdaq under “ANT”, which TestingCatalog dates to 26 October.

Prediction is not the useful part. The useful part is knowing which of your signals are worth re-budgeting on.

If you are the one deploying

1. Nothing here is purchasable, so nothing here changes your bill. No changelog line, no SKU, no price, no benchmark card. Treat a leak as marketing staging until a catalogue row appears — that row, not the screenshot, is what makes a model a product.
2. Diff your gateway’s price list in CI. LiteLLM’s file is consumed by routers that bill from it, and the 5.5 row sat there for 28 hours with a deprecation date for a model that did not exist. Pin a version of the cost map, and fail a build when a new key appears with no upstream source you can open.
3. Log the model that answered, not the model you asked for. OpenAI’s Responses API returns a model field; pin a dated ID rather than *-latest, and keep a fixed test set so a silent swap shows up as a diff instead of a vibe. The 19.8% screenshot is not interesting because a new model exists; it is interesting because the picker never told the user which build served them.
4. Do not download name-alike weights. A repo called Claude-Fable-5.5 is not Claude Fable 5.5; in this case it is a Wan 2.2-tagged GGUF. Check library_name, base_model and the file list, all of which are in the API before you pull 9 GB.
5. Watch the three things that would change this story: a line in OpenAI’s changelog, a purchasable fable-5-5/mythos-5-5 SKU in Foundry, Bedrock, Vertex or OpenRouter, and an independent benchmark entry. Each of those appeared for 5.1 first, in that order of usefulness.

The same habit is the point of two other stories on this site: Kimi K3 becoming real when Bedrock printed a price, and Mistral’s help centre still describing a licence its newest weights no longer use. Records beat announcements — including leaks.