GPT-6.2-Astra enters gray test for some Pro users as Claude Mythos 5.5 appears in site code

GPT-6.2-Astra has begun a gray test (a limited, staged rollout): the model now appears in the model lists of some users, and it works. Claude Mythos 5.5 surfaced the same day in the front-end code of the Claude website. Chinese-language outlet QbitAI (量子位) reported both developments on November 2, 2026, citing posts on X.

GPT-6.2-Astra: limited access

According to QbitAI, GPT-6.2-Astra shows up in the model list for only some users, and the model is functional when selected. Users report different routing rates:

• One user says 20% of their Astra requests were routed to 6.2.
• Another user says 100% of their requests were sent to 6.2.

QbitAI says access is mostly limited to users on the $500 Pro plan. No context window, pricing per 1M tokens, benchmark scores or parameter count for GPT-6.2-Astra were given.

Claude Mythos 5.5: briefly visible in site code

Claude Mythos 5.5 appeared briefly in the front-end code of the official Claude website and was removed a few hours later, QbitAI reports. According to the outlet, Mythos 5.5 will be the version that covers biology research and cybersecurity.

Unverified weights repository

A repository on Hugging Face claims to contain leaked Mythos weights. QbitAI says it is unlikely to be genuine, but more than 5,000 people have already downloaded it.

Pelican test comparison

QbitAI notes that both new models were put through the pelican-riding-a-bicycle prompt. The outlet's reading is that Claude emphasizes artistic style while GPT emphasizes physical realism.

What this means

GPT-6.2-Astra is in a staged test rather than a general release. The user reports show that exposure differs from account to account: one user sees 20% of Astra requests routed to 6.2, another sees 100%. Since access is mostly tied to the $500 Pro plan, the earliest hands-on feedback will come from that tier.

The Mythos 5.5 sighting is limited to a short-lived appearance in front-end code, and the described biology and cybersecurity focus comes from QbitAI's report. The Hugging Face repository should be treated with caution: QbitAI considers it unlikely to be real, and its download count reflects interest, not authenticity. Anyone considering downloading unverified weights should weigh that risk.