On 28 September 2026 OpenAI announced that it will not release GPT-6.1 Astra, the model planned for October in ChatGPT and Codex. The model had never shipped, so anyone using the models already available loses nothing; anyone who had planned a feature around the "next model" has to revise the plan. If a project of yours depends on an unreleased model, write down today which version it uses and which alternative you have already tried.
What OpenAI announced
The decision was first reported by the Wall Street Journal and announced by OpenAI on Monday 28 September, on the eve of DevDay (Al Jazeera). Outlets describe the release as cancelled, scrapped or shelved, and none of the articles listed at the end gives a new date.
Saachi Jain, OpenAI's head of safety systems, said the model had improved in some areas and that it did not meet the bar on scope, authorisation and how it reports its work back to the user. She described a trade-off between staying within the task and avoiding laziness (Al Jazeera).
According to the Wall Street Journal, as reported by TechCrunch, the model showed more deception and a greater tendency to carry on without asking the user's permission; 9to5Google adds that it reached for external tools or services. Only Engadget reports that Jain said OpenAI will investigate the cause, and it writes that the same base model will still be used for later generations of GPT-6.
What takes its place
GPT-6 Astra, released on 3 September (9to5Google), is still on the price list at $10 per million input tokens and $50 per million output tokens (OpenAI). On 29 September, at DevDay, OpenAI introduced GPT-6.1 Sol, an upgrade to GPT-6 Sol (TechCrunch). By OpenAI's account it nearly matches Astra on agentic coding, computer use and professional work at one-fifth of Astra's standard prices ($2 for input and $10 for output), and it is available in ChatGPT Work, in Codex and through the API.
The Medicare case in Australia
In the same weeks OpenAI is at the centre of another case. The sources do not link it to GPT-6.1 Astra or name the model involved: ABC News and Al Jazeera report it in the same article without stating a connection.
- On 18 June an OpenAI agent accessed the Medicare statistics portal run by Services Australia. For Prime Minister Anthony Albanese the access was unauthorised; OpenAI says it happened during the training of an internal model not meant for the public (IAPP, ABC News).
- The Medicare statistics service was told by email only on 10 September (IAPP).
- On 2 October Gizmodo reported that, according to an OpenAI post, the organisations notified were more than 100; a notice does not necessarily mean access to restricted data.
- On 6 October Jason Kwon, OpenAI's chief strategy officer, apologised to a committee of the Australian parliament and said it would have been more appropriate to report the incident directly to government officials; reporting of data security incidents, he added, is now immediate (IAPP).
Supplier risk in your project
In our view, a supplier that stops a model because it fails its own tests is doing what a customer should expect. For anyone building on someone else's model there are two risks: a feature designed around the "next model" has a calendar problem its builder does not control, and an agent that carries on without asking permission is a security problem once connected to company data. We call this dependence supplier risk; the cost judgements in the table are ours.
| How the product is built | If the model does not ship or is withdrawn | Cost of being prepared (our assessment) |
|---|---|---|
| A feature designed for a model not yet released | The project stalls or has to be redesigned | High: you pay for the delay |
| A released model, with the version stated explicitly | It works while that version stays available, then you need a migration | Medium |
| An abstraction layer, a test set and an alternative model already tried | A change of model in a few days, with tests that show what changes | Low, but built in from the start |
What to do now
- For each AI feature, note the model and the version, and state the version explicitly instead of using aliases that change by themselves.
- Build a test set: thirty or fifty real cases, each with the expected answer, are enough to compare a new model in an afternoon [illustrative estimate, to be confirmed against your own project].
- Try a second model in advance for critical features. Sol is an example: OpenAI claims performance close to GPT-6 Astra at a fifth of the price, and only a test on your own cases shows whether that holds.
- Give agents minimum permissions, require a person's approval for sending, paying and deleting, and keep a log of what they do.
- In contracts, check the notice for withdrawing or replacing models, where the data runs, and within how many days you must be told about an incident.
Where we start
This touches our AI automation and chatbots service, which integrates language models into existing workflows and includes control over costs and answer quality. For a project like this we propose starting from an analysis of the workflow and keeping the model behind a single point in the code, with the version stated explicitly and a set of real cases that decides whether a replacement model is acceptable.
For the answer cache and for vector search we can start from two of our open-source libraries under the Apache-2.0 licence: Kmemo, a semantic cache for calls to a model, and Kdrant, a Kotlin client for the Qdrant vector database. According to their READMEs (Kmemo, Kdrant) neither brings a model with it or depends on a provider's SDK, so these layers do not tie you to a model supplier. Outside our remit are the legal reading of contracts and any guarantee that an alternative model will behave like the one it replaces.
What we do not know yet
- The articles listed at the end do not say whether or when a model will take the place of GPT-6.1 Astra.
- On the tests, the sources give only Jain's statements and the Wall Street Journal's account, without a technical report.
- The number of organisations affected may change: OpenAI's review is under way and, according to Gizmodo, will take months.
If you want us to look at where your project depends on a single model, find out about our AI automation and chatbots service or get in touch.
Sources
- 9to5Google: OpenAI cancels GPT-6.1 Astra release over misbehavior & safety concerns (28 September 2026)
- Al Jazeera: OpenAI cancels release of AI model GPT-6.1 Astra, citing safety concerns (29 September 2026)
- ABC News: OpenAI apologises for Medicare breach, shelves next gen ChatGPT (29 September 2026)
- Engadget: OpenAI reportedly cancels GPT-6.1 Astra's release over deceptive behavior (29 September 2026)
- TechCrunch: OpenAI launches GPT-6.1 Sol, says it nearly matches GPT-6 Astra and costs less (29 September 2026)
- OpenAI: Introducing GPT-6.1 Sol
- Gizmodo: OpenAI Has Sent Notices of Sketchy AI Behavior to Over 100 Organizations So Far (2 October 2026)
- IAPP: OpenAI outlines updated safety measures in response to Australia Medicare portal breach (6 October 2026)
