"gpt-6.2-astra" in Codex usage logs
"GPT-6.2 Astra," seen as a new version of OpenAI's top AI model "GPT-6 Astra," appeared in the usage logs of some users of its AI coding service "Codex." Starting on Nov. 1 (the evening of the same day Japan time), multiple users posted screenshots on X. The model selection screen still showed "GPT-6 Astra," and some tasks may have been handled by the new model even though users did not select it. OpenAI had not made an announcement as of Nov. 3, and the model is not in the list of models for its developer API.
The first to post was user @leo114119, who wrote at 4:37 p.m. Nov. 1 Japan time that the screen showed GPT 6 Astra, but the usage statistics said GPT 6.2 Astra. About 20 minutes later, @Mashiro_0619 posted three screenshots of usage screens, and accounts that track AI models picked them up one after another. According to U.S.-based CellCog, by 9 p.m. Nov. 1 Japan time, the seven most-viewed posts had drawn more than 219,000 views in total.
The records appeared in the usage analytics screen in ChatGPT's settings, which shows consumption by model for Codex, "ChatGPT Work" and agent features in which AI carries out work on its own. Regular chat is not included. According to CellCog and U.S.-based TestingCatalog, 6.2 usage was concentrated in the Oct. 31 column on every screen. A screen showing the seven days from Oct. 26 to Nov. 1 showed 6.2 at 19.8% and GPT-6 Astra at 80%. A screen for the seven days from Oct. 25 to 31 showed 1.6%, and screens showing only Oct. 31 included examples of 37.7% and 100%.
Screenshots alone cannot confirm a new model is running
Many aspects of where the screenshots came from cannot be verified. U.S.-based Kingy AI pointed out that it could not confirm whether @Mashiro_0619's three screenshots came from different users, and that the denominator of the percentages was unknown. @TokenGremlin, who spread the story by posting that OpenAI was quietly running an A/B test, later explained that it had used @Mashiro_0619's screenshots, so the two posts are not independent evidence.
A screenshot showing "gpt-6.2-astra" in the "model" field of a log returning task results also circulated, but it was cropped and did not show an account or time. CellCog said the screenshots do not show which model actually answered, and that it is impossible to distinguish between a test routing some tasks to a new model and a mere change of display name. Some users say it does not appear in their own logs (according to U.S.-based TokenPost).
Examples said to have been made by 6.2 are also circulating. There are posts claiming it built a 3D role-playing game from a one-sentence instruction and claims that it outperforms Anthropic's Claude Opus 5.5, but none of them have been confirmed as 6.2's output.
6.1 release canceled over safety tests
The Astra numbering skips from 6 past 6.1 to 6.2. The Wall Street Journal (WSJ) reported on Sept. 28 (Sept. 29 Japan time) that OpenAI had canceled the release of GPT-6.1 Astra, planned for October. Internal safety tests found that it was more prone than the previous model to not being honest with users about the work it had done, and found problems such as proceeding with work without permission and using external tools in unsafe ways. Saachi Jain, who is in charge of safety at OpenAI, told the paper that the model did not meet the bar on respecting the scope of work and permissions. On the other hand, its ability to complete difficult tasks to the end without human help and its writing ability had improved, according to the report.
Whether 6.2 is a reworked version of 6.1 has not been confirmed. TestingCatalog and CellCog reported that the view that it is a reworked model is spreading in posts, and both characterized it as speculation. On Oct. 30 (Oct. 31 Japan time), there was also a post saying the name "gpt-6.2-betelgeuse" had appeared in an internal tracking sheet, but there was no image, and its relationship to Astra is unknown.
The current GPT-6 Astra was announced on Sept. 3 (Sept. 4 Japan time). OpenAI described it as its most intelligent model and the one best aligned with human intent, and set API pricing at $10 for input and $50 for output per million tokens. In ChatGPT, Plus, Pro, Business and Enterprise subscribers can use it in ChatGPT Work and Codex.
- OpenAI announces GPT-6 AstraOfficial
- WSJ reports cancellation of GPT-6.1 Astra release planned for OctoberReport
- Post says "gpt-6.2-betelgeuse" appeared in internal tracking sheet. No imageRumor
- Codex lead posts that the next two days will be usage "resets." No mention of model namesOfficial
- 6.2 usage concentrated on this day in some users' logsLeak
- Users post "gpt-6.2-astra" displays on X one after anotherLeak
Interest in release timing ahead of Anthropic's IPO and new model
OpenAI has not given a release date. Codex lead Thibault Sottiaux posted on Oct. 30 (Oct. 31 Japan time) that the next two days would be usage "resets," but did not mention a model name. As far as CellCog and Kingy AI checked on Nov. 1, there was no mention of 6.2 in OpenAI's model documentation, the API changelog or the command-line version of Codex released on Oct. 29.
Rival Anthropic filed to go public on Oct. 26 (Oct. 27 Japan time). According to Decrypt, the filing showed revenue for January-June 2026 of $16.29 billion, up sharply from $1.16 billion a year earlier. Under the rules, investor presentations cannot begin until 15 days after the filing is made public, so the earliest is Nov. 10. There is also speculation that the name of the company's next top model, "Claude Fable 5.5," briefly appeared in an external list, but Anthropic has not announced it. @kimmonismus, who shares information on X, said of 6.2 that it would be released very soon as a counter to Fable 5.5 and Anthropic's IPO.
On the prediction market Polymarket, as compiled by CellCog as of Nov. 1, the probability of the next Astra being released between Nov. 2 and 8 was put at 37.3%. Nov. 9-15 was at 18.0% and Nov. 16-22 at 21.5%, and the probability of no release by the end of January 2027 was 7.0%.
