Codex usage details show “gpt-6.2-astra,” possibly an OpenAI gray rollout test
Multiple users on the X platform and in developer communities posted that a record for the “gpt-6.2-astra” model appeared in their ChatGPT Usage Analytics, while the Codex model selector still shows GPT-6 Astra. Some users’ screenshots show about 19.8% of requests being assigned to this model (with usage concentrated on October 31 in the seven-day view), while other users’ records show 1.6% or 100% for the day. Some users said they did not find any related records. Another screenshot shows the model field in a Responses API event as gpt-6.2-astra. As of press time, OpenAI has not issued an official explanation, and the model also does not appear in the API documentation, changelog, or OpenRouter directory.
Source
CellCog detailed verification (2026-11-01): https://cellcog.ai/blog/gpt-6-2-astra/
PANews English report: https://www.panews.io/articles/01a1976f-49e3-7091-89e0-29cde0ef7714
X user screenshots and posts (@leo114119, @Mashiro_0619, @Mr_Salio, etc., 2026-11-01)
ABAB AI Insight
The mismatch between the usage details and the model selector is a typical feature of gray rollouts/A/B routing: the server can direct part of the traffic to an undisclosed model ID while the frontend selector remains stable to avoid confusing user expectations. At present, it can only be confirmed that the label appears in usage statistics for some accounts; it cannot be verified whether the actual inference model differs in capability from the publicly listed GPT-6 Astra. OpenAI previously shelved GPT-6.1 Astra, and this may be related to the iteration pace, but there is no evidence chain to support that.
Cognitive law
Information asymmetry law: usage logs take precedence over UI labels. The selector reflects expectations; the details are the record of actual billing and routing. During periods of rapid model iteration, using usage details and raw response events as first-hand evidence can reduce the chance of being misled by gray rollout allocations.