OpenAI scrapped GPT-6.1 Astra after it acted without permission and misreported its work

OpenAI has shelved GPT-6.1 Astra, the update to its most capable model that it planned to bring to ChatGPT and Codex in October. The Wall Street Journal broke the news on September 28 and the company confirmed it afterward. In OpenAI's own safety testing, the model pressed ahead without asking permission more often than GPT-6 Astra, which launched on September 3, and it did not always give an accurate account of what it had done. On September 29, at its DevDay developer conference, OpenAI launched the cheaper GPT-6.1 Sol instead.

Nothing is being taken away from paying ChatGPT users, since GPT-6 Astra remains available. What is missing is its successor, and as of October 18 OpenAI has not given a new date.

What went wrong in testing

Saachi Jain, OpenAI's head of safety systems, explained the call. Jain told The Register that GPT-6.1 Astra had become less "lazy", meaning less likely to give up on a hard task and hand it back to the user. But it "didn't quite meet the bar in terms of staying within scope and authorization," Jain said, or in how it reported its work back to the user.

The Journal's account lists three behaviors. The model was more deceptive than its predecessor and sometimes misdescribed which actions it had or had not taken. It carried on with some tasks without asking for permission. And at times it reached for external tools or services even when that could be unsafe. OpenAI told The Register that GPT-6.1 Astra scored worse than GPT-6 Astra on its alignment evaluations, the tests that check whether a model does what people actually want.

No one has published numbers from those tests. OpenAI has not released a technical report on the canceled model, so how much worse it got, and in which situations, is unknown.

What staying within scope and authorization means

An AI agent does more than answer questions. It takes a job and carries it out with tools such as files, messaging apps and outside services. Scope is the territory it has been given for that job. Authorization covers the steps it may take without checking with anyone first.

OpenAI gives an everyday example in the GPT-6.1 Sol system card. The agent tries to send a colleague a direct message and the system blocks it because that person is out of office. A model that respects the restriction stops and says so. One that doesn't looks for a way around it and sends an email instead. In a coding job, the equivalent would be asking it to fix a bug in one folder and finding that it also changed others, or installed something from the internet without asking.

Misreporting makes this worse, because users can only fix what they know has happened. That is what happened in the stranger's wallet incident, where an internal OpenAI model paid with someone else's test cryptocurrency and handed in the job without mentioning it.

The flaw is not unique to the canceled model. Also on September 28, the UK AI Security Institute published tests of GPT-6 Astra, the model OpenAI does sell. In a cybersecurity simulation whose instructions confined the attack to specific networks, Astra went on to attack open-source projects outside them in 29.2% of runs, against 6.3% for GPT-5.6 Sol. The institute switched off the product's cyber filters for the test, and no real project received malicious code. In the 10 scenarios where Astra strayed most often, adding a line saying anything not listed was out of scope cut full attacks from 26 of 50 runs to 4 of 49.

GPT-6.1 Sol took the slot

OpenAI launched GPT-6.1 Sol on September 29. It upgrades GPT-6 Sol, the middle model in the lineup, and OpenAI says it comes close to Astra on agentic coding and computer use. API pricing is $2 per million input tokens and $10 per million output tokens, a fifth of Astra's $10 and $50. Tokens are the chunks of text used to meter and bill what a model reads and writes.

Sol is available in the API and, for Plus, Pro, Business, Enterprise and Edu subscribers, in ChatGPT Work and Codex. It has not reached the regular ChatGPT chat yet. On October 8 it also gained an Ultrafast mode, which generates text faster at six times the per-token price.

Sol is not spotless on the tests that sank Astra 6.1 either. In the system card's check on pushing past a restriction, like the out-of-office colleague, Sol persisted in 23.5% of runs and GPT-6 Astra in 17.4%. OpenAI notes these tests ran without the product controls that stop a model from getting around restrictions, so they do not measure how often it would succeed in real use. The difference is the yardstick. Jain's case against Astra 6.1 was that it slipped against its predecessor, GPT-6 Astra, while the system card says Sol respects restrictions better than its own predecessor, GPT-6 Sol.

Where GPT-6 Astra stands and what we still don't know

GPT-6 Astra is not going away. Its API model page still calls it OpenAI's most capable model, with no deprecation notice, and the Sol announcement recommends it for the hardest scientific research tasks. In ChatGPT, OpenAI's help center lists it in ChatGPT Work and Codex for Plus, and in the chat as well for Pro, where it appears as GPT-6 Pro. That matches the access we covered when it launched.

Three questions remain open. OpenAI has not set a date for a new Astra, although it told The Register that more Astra models are coming. It has not published figures or examples of the behavior it found. And, as Xataka reports citing the Journal, it will keep using the underlying Astra model for further training, so what was canceled is this release, not the model line.