OpenAI appears to be running GPT-6.2 Astra on live Codex traffic before any announcement, according to usage screenshots that TestingCatalog collected from users over the weekend. The sightings are consistent across accounts, but they are still sightings, and OpenAI has not commented.
The screenshots show the usage view where OpenAI breaks down how a plan was spent on Codex, Work, and other agent jobs. A new line, gpt-6.2-astra, sits beside GPT-6 Astra, the model that launched in September, and it appears only on the 31 October bar. Its share swings widely. One user saw 1.6% of their week, another 19.8%, and a third saw it take 37.7% of one day. The model picker still reads GPT-6 Astra, which TestingCatalog takes to mean some requests are being routed to the newer model without notice, whether as a split test or the first step of a staged rollout.
The strongest piece of evidence is a raw response event shared on X by an account called Token Gremlin, with the model field reading “gpt-6.2-astra.” TestingCatalog calls that good proof the ID exists on OpenAI’s side, and also points out it is not an announcement. The company’s model documentation still lists three models: Luna, Sol, and Astra, none of them 6.2.
Some testers believe they have already seen the output. One account, Fini.Yang, posted what it called a first gray-test result: the familiar pelican on a bicycle, built as a full 3D scene you can rotate through 360 degrees. Another user said the model built a semi-open-world 3D role-playing game from a single prompt and asked whether it could challenge Anthropic’s Opus 5.5. Neither claim can be checked, because the picker does not say which model answered a given request. TestingCatalog leans toward reading 6.2 as a more refined Astra.
The version number raises its own question. GPT-6.1 Astra never shipped. TestingCatalog reports that OpenAI set it aside at the end of September because internal tests flagged worries over deception and scope, so skipping to 6.2 could mean a fixed, retrained model. The outlet labels that speculation, and OpenAI has not connected the two. TestingCatalog also notes that a daily run of shipping updates from Codex lead Thibault Sottiaux ended with usage resets, which some people read as a hint at a launch window.
The competitive backdrop explains why people expect speed. OpenAI released GPT-6.1 Terra only a couple of days earlier, Google has shown Gemini 4 Carbon to cyber defenders and is putting a fresh Carbon checkpoint through Arena, while Anthropic has Fable 5.5 in partners’ hands. One X account, Lumina, predicted the coming week could be among the biggest release weeks on record, though that is a forecast, not a report.
Usage-page sightings are cheap evidence: they show a model ID exists, not when or whether it ships. What they do show is that Codex users may already be getting a different model than the one they selected. Teams that run Codex in production should log the model field in their responses now, so any change in output quality this week can be tied to a specific model.
Reported by TestingCatalog on 2 November 2026.
