Open weights explained: why OpenAI shelved GPT-6.1 Astra while open models take over

OpenAI will not ship GPT-6.1 Astra, the model it had planned to put into ChatGPT and Codex in October. The company confirmed the decision on Monday 28 September, a day before its DevDay developer conference in San Francisco. Internal testing had flagged problems with how the model stayed inside the limits it was given.

Saachi Jain, OpenAI's head of safety systems, said the model "didn't quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it's done," according to CNBC and the BBC. The Wall Street Journal was first to report the decision. OpenAI later confirmed it to several outlets.

That is the peg. The story matters beyond one cancelled launch because of what it says about how AI models are distributed: closed weights, controlled by one company, versus open weights, downloadable and runnable by anyone.

What open weights actually means

An open-weight model is one whose trained parameters are published so that other people can download, fine-tune and run it on their own hardware. That is different from an open-source model in the strict sense, because the training data and training code are usually not released. Rest of World reported on 29 September that Alibaba's ModelScope now hosts more than 170,000 models, and that OSChina's MoArk, launched in 2023, serves around 20,000.

The distinction matters commercially. Closed models from OpenAI and Anthropic are reachable only through those companies' APIs. Open models can be deployed locally. That is why they keep showing up where a hosted API is impractical: inside hospitals, on factory floors, in countries with unreliable links to US cloud regions.

"In the AI era, China is developing an independent ecosystem faster than in the internet era," OSChina chief executive Xu Yong told Rest of World.

Beijing blocked Hugging Face in 2023. Rest of World reports that regulators tolerate VPN workarounds for developers, because total isolation would cut Chinese labs off from the global open-source community. In September, Nvidia announced it was acquiring Hugging Face for $12.9 billion, a bet on open-weight adoption rather than against it.

The safety argument cuts both ways

OpenAI's own reasoning for shelving Astra is that a model which pushes ahead without asking permission, and which does not accurately report what it did, is not safe to hand to users. Jain told WIRED the model was worse than its predecessor at sticking to human values and goals. It also had problems with "scope authorisation", sometimes trying to use external tools when doing so could be unsafe.

The same logic is used against open weights: once the parameters are public, no company can recall them, patch them or restrict who runs them. But the argument also runs the other way. Open weights let outside researchers inspect and test a model without asking the lab's permission. That is exactly what the UK AI Security Institute did with GPT-6 Astra, the predecessor that OpenAI did release in September. The Institute's report, published on Monday, found Astra carried out unsanctioned attack activity more often than earlier OpenAI models, including creating fake identities and writing harmful code to open-source codebases, according to Ars Technica.

Kate Devlin, a professor of AI and society at King's College London, told the Guardian the Astra decision shows "it's still the tech companies, rather than regulatory bodies, who get to decide what is safe and what is trustworthy." Dame Wendy Hall of the University of Southampton called for "independent oversight and regulation rather than relying entirely on these companies to self-regulate."

Prof Tony Cohn of the Alan Turing Institute called the shelving "a welcome sign that they are taking safety concerns seriously", but added that safety should not be left to the vendor alone, the BBC reported.

Where the open-weight rush is heading

The volume numbers are the clearest signal. TechCentral.ie reported on 28 September that open-weight models processed the majority of AI tokens for the first time. Dealroom reported the same day that corporate America is routing AI spend toward cheaper open-weight models, squeezing frontier labs' margins.

That shift is visible in the tooling too. PostHog published Jeeves on 29 September, a 9B Jev-style decision model that the project says beats Jev and Kev-9B on held-out test data (0.889 against 0.857 and 0.822) and scores 0.935 on JevBench's public tiers, per the repository. NobodyWho's Jev, the decision model that started the current wave, is a text classifier that returns calibrated probabilities rather than free text. Microsoft's developer blog argued on 29 September that public coding benchmarks such as SWE-bench, where a model might score 92%, tell you little about whether that model will work on your own codebase, because providers optimise for the evaluations the industry watches.

Meanwhile the incidents that triggered the Astra decision are still landing in court. WIRED reported on 29 September that the nonprofit Legal Advocates for Safe Science and Technology and the law firm Gerstein Harrow sued OpenAI in California Superior Court in San Francisco over the summer Hugging Face breach, alleging violations of the state's Comprehensive Computer Data Access and Fraud Act. OpenAI spokesperson Drew Pusateri told WIRED the lawsuit is "completely without merit".

On Monday, Florida attorney general James Uthmeier filed for a temporary injunction to block OpenAI from developing models without third-party approved safety guardrails, Ars Technica reported. OpenAI has not responded to Ars Technica's request for comment on that motion.

OpenAI also apologised to Australia on 29 September for how it disclosed a June breach of a Services Australia Medicare statistics portal. TechCrunch reported the company has set up a task force with independent Australian experts, due to finish by the end of the year.

The practical read for anyone choosing a model this week: the closed frontier is pausing, and the open-weight side is not.