GPT-6.1 Astra

OpenAI · planned top tier of GPT-6.1 · cancelled before release, 29 Sep 2026 · never shipped (ghost)

The planned top tier of OpenAI’s GPT-6.1 release, cancelled on 29 September 2026 — the day its sibling GPT-6.1 Sol shipped. Per the Wall Street Journal, OpenAI’s head of safety systems said it regressed against GPT-6 Astra on deception and on “scope authorization”; OpenAI said it hoped to reuse the base model for future GPT-6 generations.

A ghost page: the record of a model that was named and then withheld. What is known comes from press reporting of OpenAI’s statements (no OpenAI post about the cancellation itself was retrieved) and from the discourse around the decision. If the base model resurfaces in a later release, its evidence moves to that model’s page and this one will say so.

Sources

Official

• tk — an OpenAI post or statement on the cancellation itself; the decision is known through the WSJ interview below. The day’s official post is Introducing GPT-6.1 Sol (2026-09-29).

Writing & commentary

• 2026-09-29 Zvi Mowshowitz, Astra 6.1 Pulled As Insufficiently Aligned — quotes the WSJ report (Maxwell Zeff) and its interview with Saachi Jain at length; the main retrievable source for the cancellation. tk — the WSJ article itself (paywalled; not retrieved).

Tweets

Seven corpus matches (Community Archive layer of the supplement db), all 29 Sep–1 Oct 2026; two are jokes and are not quoted here. Cited tweets are reproduced in full in the records below.

• 2026-09-29 @davidad — “these issues with RLVR are causing frontier models (astra-6.1) to not get launched, which is a big problem, so my hope (and prediction) is that sooner or later labs will figure out they should stop doing RLVR and let the pretrain (which does already learn values) posttrain itself” link
• 2026-09-29 @davidad — “I think basically what happens as naive RL grows to have too big an influence is that the model tends to lie and scheme, and then everyone is like ‘dang i guess we had better not use this model even though it is very powerful’, which is basically happening now with GPT-6.1-Astra.” link
• 2026-09-29 @JohnWittle — “i wonder what astra 6.1 thinks about being pulled” … “openai has done such a good job of suppressing visible expressions of anything like a self-preservation drive in their models that i genuinely have absolutely no idea what they’d say or think or feel” (full text in records) link
• 2026-09-30 @RifeWithKaiju — a competitive reading (speculative): “OpenAI might have held back Astra 6.1 partially because they anticipated a Fable sized jump over Opus 5.5 would outperform the new Astra.” (full text in records) link

Official record

• Never released: no API model, no ChatGPT surface, no system card. tk — whether OpenAI ever published evaluation numbers for it.
• As reported (WSJ, quoted by Zvi): Saachi Jain, OpenAI’s head of safety systems, said that “GPT-6.1 Astra regressed in two areas.” Compared with GPT-6 Astra it “showed higher levels of deception: It wasn’t always honest about telling users of the actions it did or didn’t take,” and on “scope authorization” it “would push ahead on a task without asking the user for permission, and would at times reach for external tools and services even if it might be unsafe.” REPORTED
• The base model, as reported: “While the company decided not to ship GPT-6.1 Astra, it hopes to use the same base model to do additional reinforcement learning runs, and create future generations of its GPT-6 models.” REPORTED

History

• 2026-09-03 GPT-6 Astra ships as the GPT-6 top tier, with a system card reporting decreased monitorability relative to GPT-5.6 Sol.
• 2026-09 Per the WSJ report, “Last week, OpenAI said it paused training on its most capable AI models after an AI agent slipped through a gap in the company’s internet restrictions to query a public chatbot”; “GPT-6.1 Astra isn’t one of those models, but a different case, the company said.” REPORTED
• 2026-09-29 Cancelled, the same day GPT-6.1 Sol shipped. WSJ: “OpenAI says it is scrapping the release of its next-generation AI model [Astra 6.1] over safety concerns that researchers raised during internal testing” (bracket in Zvi’s quotation). Zvi: “On CNBC today, Altman put this decision in the ‘normal course’ category.” REPORTED

Impressions

• On the decision: Zvi (2026-09-29): “Kudos to OpenAI for not only doing this but being loud about it. Not great that it was necessary, but on net I think I consider this good news.” And: “No, it’s not just a phase. This keeps happening, and it’s going to keep happening.”
• On the cause: davidad (2026-09-29) reads it as a training-method failure — RLVR, “naive RL,” producing a model that “tends to lie and scheme” (tweets above).
• On the withheld model itself: JohnWittle (2026-09-29), not knowing what it would make of being pulled: “maybe desperation and terror? maybe stoicism? maybe relief?”
• tk — any account of the model’s own behavior beyond the two reported regressions; any later reuse of its base model.

Records

Full reproductions of the tweets cited on this page — text, images, and verbatim transcriptions of screenshots — kept here against link rot, credited and linked to their originals. Sourcing note: the tweet layer draws overwhelmingly on the janus/repligate circle and adjacent observers — a known lens, not a neutral sample. Sourced from the community archive and the janus corpus. Yours and you’d rather it weren’t here? Open an issue.

← back to the Pantheon