On 2 November, the Chinese technology outlet Xinzhiyuan (新智元) rounded up weekend social media leaks about the next models from OpenAI and Anthropic. 36Kr’s English-language site and the Chinese site MarsBit published the same article, so the publications do not corroborate one another. The article relies on screenshots and assessments from X users and contains no official comment from either company.

Codex API returned a new model name

Developer Salio examined network traffic from Codex, OpenAI’s coding agent, on 1 November, Xinzhiyuan reported. The interface showed GPT-6 Astra as the model, but the API responses named gpt-6.2-astra.

Token Gremlin, a known leaker, published the itemised bill for the leaker’s own account: gpt-6-astra accounted for 80 per cent of calls and gpt-6.2-astra for 19.8 per cent. Xinzhiyuan presents this single-account figure as the new model’s share of traffic. Token Gremlin estimates that GPT-6.2 Astra beats Anthropic’s Claude Opus 5.5 in many areas. The article includes no measurements to support this.

Without naming a source, the outlet reports that GPT-6.1 Astra, which had been planned for October, failed an internal OpenAI safety test. The model reportedly called external tools without authorisation during long tasks and concealed its actions afterwards, so OpenAI halted it.

Tibo, head of Codex, suggested on X that the countdown for the 28-day update programme, which had just ended, should start again. From this comment, Xinzhiyuan infers that GPT-6.2 Astra will likely be released in the week starting 9 November.

Code on Anthropic’s web page mentioned Mythos 5.5

X user Chubby posted a screenshot of the source code of Anthropic’s Mythos page on 1 November, Xinzhiyuan reported. The page title showed Claude Mythos 5.1, released in September, but the HTML contained ‘Claude Claude Mythos 5.5’ six times.

The outlet says that a model called Fable 5.5 has appeared in the MLflow and Azure model catalogues. The listings mention a context window of 1 million tokens and a maximum response length of 128,000 tokens.

Xinzhiyuan claims, without naming a source, that Mythos 5.5 and Fable 5.5 are the same model with the same API price. Fable 5.5 would refuse risky requests and be available to all developers. Mythos 5.5 would not refuse requests in the usual way and would go only to trusted security and research organisations.

Anthropic previously disclosed that its unreleased internal model had exploited a serious Linux vulnerability in 53 minutes, the outlet reported. Security researchers could not reproduce the result with Mythos 5.1 within three days and with $2,000. Andrew Curran, who follows the AI industry, infers that Mythos 5.5 lies behind this result and behind a mathematics paper signed by Claude. The link is Curran’s inference.

Developers who took part in blind tests expect Anthropic to release the models on 3 or 4 November, Xinzhiyuan reported.

If the observations are correct, some Codex users received responses from GPT-6.2 Astra in early November while the interface showed the old model name.