Claude Mythos 5.5 and GPT-6.2 traces: what's real, and the fake 'Numbat' leak

Three things went viral on Sunday. First, @kimmonismus shared a screenshot of Anthropic's Mythos web page whose hidden code read 'Claude Mythos 5.5'; independent checks by Hum and FourWeekMBA confirmed that a hidden heading and the logo's alt text said 5.5 on Sunday afternoon before being changed to 5.1, while the visible page always said Mythos 5.1. Second, a model called 'gpt-6.2-astra' showed up in some users' Codex usage breakdowns (spotted by @TokenGremlin and shared as 'Confirmed. GPT-6.2 very close' by @kimmonismus); the model picker still says GPT-6 Astra and asking Codex for gpt-6.2-astra returns 'model not found'. Third, a post with 288,000 views claimed a '14-trillion-parameter Mythos-class' Anthropic model called 'Numbat' had leaked on Hugging Face. @kimmonismus called it likely fake, and the forensic write-ups agree: the files are seeded random numbers with no link to Anthropic. Fact or hype: partly fact. The traces are real and suggest both labs have new versions close, but neither Anthropic nor OpenAI has announced anything, so 'confirmed' is overstated. The Numbat leak is hype: it's the second dubious 'Anthropic weights' story in two days, after the 'Fable 5' claim in yesterday's edition.

Why it mattersDon't change plans or budgets on a page-source screenshot; wait for the release notes. And never download 'leaked' model files or run their custom code: at best they're junk, at worst they're malware.

Curated by Sara Simeone as part of our human editorial coverage. The original sources are linked below; curation is distinct from a Hype or Fact? assessment.

How we knowCompany figures are labelled as their own claims; anything we couldn't verify is labelled as a report or a claim.