Anthropic's Website Briefly Carried "Claude Mythos 5.5" in Its HTML, and a Guarded "Fable 5.5" Is Already in Model Catalogs

Anthropic's own website briefly advertised a model that does not officially exist. For a few hours around November 1, the HTML source of the Claude Mythos page contained six references to "Claude Mythos 5.5" — including one accidental doubling, "Claude Claude Mythos 5.5" — before being reverted. The page currently displays Mythos 5.1, and Anthropic has made no announcement. We checked the page again on Sunday: no trace of 5.5 remains in the production markup.

The sighting landed in the same weekend that OpenAI's unannounced GPT-6.2 Astra surfaced in Codex billing data, which is to say: both leading labs appear to be weeks from flagship releases, and both are leaking through infrastructure rather than press releases.

Chinese tech outlets piecing the signal together describe a two-tier split of the same underlying model. Fable 5.5 would be the guardrailed public flagship; Mythos 5.5 the version with refusal restrictions loosened, reserved for vetted cybersecurity teams and research institutions — reportedly at identical API pricing. Fable 5.5 has independently appeared in MLflow and Azure model-catalog listings with a 1 million-token input window and a 128,000-token output cap, though those entries are likewise unofficial and could change before launch.

Bloggers have been quick to attribute two recent unnamed-model feats to the checkpoint: the 53-minute Linux kernel exploit that Anthropic disclosed in October, and a machine-authored mathematics paper that circulated with no human author listed. Anthropic has connected neither to Mythos 5.5, and neither claim should be treated as confirmed.

A Hugging Face repository under the handle burner22/numbat also claims to be leaked Mythos weights, with a model card describing a roughly 14-trillion-parameter mixture-of-experts architecture. More than 5,000 people have downloaded it. The sensible reading, shared by most of the AI community, is that it is not genuine — but the download count says something about demand.

The commercial logic of a two-tier lineup is easy to see. Anthropic already restricts Mythos 5 to a set of US organizations following US government export-control decisions — that much is stated on its own site. And evaluations like SemiAnalysis' new GDPx benchmark show safety classifiers shaving measurable points off top models in cyber and biology, which gives any lab a reason to sell an unguarded tier behind a vetting process rather than sandbag the public one.

Nothing here is confirmed, and the history of pre-launch leaks is a history of changed names and moved dates. But if the structure is real, the interesting question is not which flagship tops the next benchmark. It is who gets the un-sandbagged one — and whether the guarded and unguarded versions ship the same week.