Not on OpenAI's Model List, but Already in Codex Billing: GPT-6.2 Astra Shows Up at $0.41 to $0.52 a Request

Over the weekend, developers started noticing that the model behind their Codex requests had quietly changed. The interface still says GPT-6 Astra, but usage statistics and billing records began returning a different identifier: gpt-6.2-astra. Multiple independent screenshots circulated on X and the Chinese developer forum Linux.do on November 1, which reads more like a broad gray rollout than a single misconfigured account.

The earliest widely shared evidence came from a Linux.do post published at 02:36 Beijing time on November 1 by a user asking whether OpenAI had started a new round of gray testing. The attached usage panel covered October 25 to 31 and showed gpt-6-astra taking 97.7 percent of requests, with the unannounced gpt-6.2-astra at 1.6 percent. The remainder split across known models: gpt-5.6-terra at 0.4 percent, gpt-6.1-sol and gpt-5.6-luna at 0.2 percent each, and gpt-6-luna below 0.1 percent. The poster said he selected Astra with extra-high thinking and standard speed, reported peak throughput around 90 tokens per second, and attached 13 billing lines — 12 marked gpt-6-astra and one marked codex-auto-review — with individual requests costing between $0.41 and $0.52 and logged speeds of 44 to 73 tokens per second.

A second, harder-looking data point came from the well-known leak watcher Token Gremlin, whose Pro account usage panel showed gpt-6.2-astra at 19.8 percent of traffic against 80.2 percent for gpt-6-astra — a very different snapshot, which either reflects a different account tier, a later window, or a fast-expanding rollout. None of these screenshots has been independently authenticated, and another shared capture showed the system first throwing an unsupported-model error before successfully completing a /review task with gpt-6.2-astra in the status bar. OpenAI's public model catalog does not list GPT-6.2 Astra, and the company has said nothing.

The version number itself is the tell. OpenAI scrapped GPT-6.1 Astra days before its DevDay conference in late September after internal tests flagged safety problems, as the company's own community reported at the time. The obvious inference — that the rebuilt model is shipping as 6.2 after roughly a month of rework — is plausible but unconfirmed, and some users argue a reworked Astra would not carry a 6.2 number at all.

The performance claims flying around the community should be treated with equal caution. Posts describe large gains over GPT-6 Astra: cheaper, faster, fewer tokens per task, noticeably better design taste and coding ability, and even superiority over Claude Opus 5.5 in nearly every area. No system card, benchmark or official statement backs any of that. The counter-evidence is also anecdotal: one tester reported an 11-minute task with no speedup at all, and a shared SVG generation contained visible clipping artifacts.

The timing is not accidental. Within hours of the OpenAI sightings, strings reading Claude Mythos 5.5 and Fable 5.5 were spotted in Anthropic's own website source code — an unofficial sighting of the same kind, and one that suggests both labs are racing their next flagships into internal traffic at the same moment. Gemini 4 Carbon claimed the top of the Artificial Analysis index only days earlier, so the weekend's dual leaks land in the middle of an unusually tense leaderboard fight.

What this episode really demonstrates is that silent gray releases have turned billing data into a disclosure channel. Developers now learn about frontier model changes from their own invoices and usage panels, not from launch posts. For enterprise buyers that is more than gossip: if the model behind a paid endpoint can change underneath a fixed name, version pinning, audit trails and procurement contracts all need explicit language about it. Until OpenAI speaks, GPT-6.2 Astra remains a sighting, not a launch — but one that is already being paid for, one request at a time.