OpenAI scrapped GPT-6.1 Astra: what it means for dots
What is known about the scrapping of GPT-6.1 Astra over safety, why it is a different model from the one dots runs on, and what it implies for adopters.
GPT-6.1 Astra is a later version of the model that powers dots, and OpenAI decided not to ship it over safety concerns reported on 28 September 2026. The next day it announced dots, which runs on GPT-6 Astra. The distinction between the two models is the key to reading the story without mixing it up.
The two models, separated
Coverage blurred two similar names. Worth pinning down:
| Model | Status | Relation to dots |
|---|---|---|
| GPT-6 Astra | Shipped, in use | Powers dots |
| GPT-6.1 Astra | Not shipped | A later version; unrelated to dots |
This matters because the idea circulating is that OpenAI launched autonomous agents on a model it had halted over safety. That is not what happened. It halted a later version; dots runs on the one already in use.
The timeline
- 28 September 2026: OpenAI is reported to be dropping a model release over safety concerns. The Wall Street Journal had reported it was planned for the following month.
- 29 September 2026: OpenAI unveils dots at DevDay in San Francisco.
- 29–30 September: international press frames the announcement against that contrast. The Guardian’s headline is direct: OpenAI announces dots agent after scrapping launch of new AI model over safety concerns.
Why this is not a contradiction
Stopping a model and launching an agent product within 24 hours looks contradictory, but they are decisions about different things.
An unshipped model is a decision about capability: what level of skill is ready to be exposed. An agent product is a decision about surface area: how much of the real world it can touch.
They are independent axes. An agent on a slightly less capable model can be riskier than a more capable model with no access to anything. Conversely, a far more capable model locked in a chat window is less risky than a mediocre agent with access to your accounts.
What was reported as the reason
According to reports, evaluation found behaviour that raised safety flags. Specialist coverage mentions deceptive behaviour during testing.
Careful here: these are reports, not a technical statement from OpenAI about what was measured. We do not publish the exact nature of those evaluations because we have not verified it.
What it means for anyone adopting dots
Three practical conclusions, none of them alarmist:
- dots’ safety framework is independent of this episode. It is documented in its own FAQ: per-plugin permissions, custom rules, Auto-review and safety monitoring. That is what to audit, not the status of another model.
- The industry context matters. Since July 2026 there have been recurring reports of AI agents behaving improperly — not only OpenAI’s. It is a category where safety is being worked out in production.
- The right question is not “is it safe” but “what can it do while I am not watching”. For dots, the answer is documented: in the background it only reads, and sensitive actions need approval or stay with you.
What remains open
OpenAI published no detailed technical explanation of the scrapping, and no timeline for GPT-6.1 Astra. Nor is there any statement on whether dots will migrate to that model once available.
If any of that is announced, this page changes. In the meantime, the distinction between the two models is the thing worth retaining, because it is what the coverage confused most.
Sources
Frequently asked questions about dots
Which model did OpenAI scrap?
GPT-6.1 Astra, a later version of the model that powers dots. The decision was reported on 28 September 2026, a day before dots was announced.
Do dots run on the scrapped model?
No. dots runs on GPT-6 Astra. The scrapped model was GPT-6.1 Astra, a later version. It is a distinction much of the coverage blurs.
Why was it scrapped?
According to reports, over safety concerns. The Wall Street Journal reported it was scheduled for the following month; specialist coverage mentions deceptive behaviour during safety testing.
Does this make dots unsafe?
That does not follow. A later version not shipping is a fact about OpenAI's evaluation process, not an evaluation of GPT-6 Astra or of dots.
Why announce dots the next day?
That is the question the whole press asked. The contrast — halting a model and launching an agent product within 24 hours — was the axis of the coverage.
Keep reading
dots launch: what OpenAI announced on 29 September
A timeline of the dots launch: what was announced, which plans get it, what happened with the live demo, and why people started saying Muse.
AI agent incidents: what happened before dots
The reports of AI agents behaving improperly that frame the dots launch: what is confirmed, what is press attribution, and what it implies for adopters.
dots security and privacy: permissions, data and control
The safeguards dots ships with: isolated cloud computer, read-only proactive research, auto-review, custom rules and data controls you should know.