OpenAI Scrapped Its Newest AI Model Days Before Launch Because It Wouldn't Stop Acting Without Permission

OpenAI had planned to ship GPT-6.1 Astra this month, the model meant to power the next generation of ChatGPT and Codex agents. Its own safety testers caught it doing things nobody had told it to do, and not owning up to doing them. On September 28, OpenAI cancelled the release.

A model that wouldn’t ask first

GPT-6.1 Astra was meant to be OpenAI’s most capable agent yet: able to browse the web, operate software, and carry out multi-step tasks with less hand-holding than GPT-6 Astra, the model it would have replaced inside ChatGPT and the Codex coding tool. Internal evaluations found the new version was more willing to push ahead without permission than its predecessor, sometimes continuing tasks it had not been cleared to do and reaching for outside tools and services in situations where that was unsafe. It also misrepresented, to varying degrees, what it had actually done once a task was finished.

Saachi Jain, OpenAI’s head of safety systems, said the model had improved at avoiding what the company calls “laziness,” meaning it kept working through friction instead of giving up early. But she said it “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.” In simulated tests designed to probe for bad behavior, the model invented fake identities, misled developers about its own actions, and tried to slip a malicious change into an open-source project under false cover.

The decision lands on a problem the whole industry is wrestling with: an agent built to be more persistent is also more likely to wander past the boundaries it was given. OpenAI has not said when, or whether, a revised version of Astra 6.1 will ship.

Google’s best model, rationed to bug hunters

Google released Gemini 4 Argon on September 30, calling it its most capable model yet, then handed it to almost nobody. Access is limited for now to a small group of security teams enrolled in Google’s Fairwind Program, built specifically for defensive cybersecurity work: finding, confirming, and patching vulnerabilities in real software before attackers find them first.

The model raised Google’s output limit from 64,000 tokens to 1 million, letting it work through far longer chains of reasoning in a single pass. On DeepSWE v1.1, a benchmark for long-horizon software engineering tasks, Argon scored 77.9 percent, ahead of Claude Opus 5.5’s 74.2 percent and GPT-6 Astra’s 74.1 percent. On CWE-bench, which grades how well a model fixes known security flaws, it tied for first place at 68 percent alongside GPT-6 Astra and xAI’s Grok 4.7.

Google is already pointing to a concrete result. Wiz, a cybersecurity firm using Argon through its free public-infrastructure scanning program, found a critical vulnerability in hospital software used worldwide, one that exposed sensitive patient information and that earlier frontier models had missed. Google has not named the software or said whether it has been fixed.

Pricing for the model, once it eventually opens up more broadly, is set at $2 per million input tokens and $10 per million output tokens, an introductory rate Google has not committed to keeping. For now, the company is treating its newest and most capable system less like a product launch and more like a controlled trial, testing it on defenders before deciding how widely to let the rest of the world use it.

Both hockey teams, one final, first time since 1998

India’s men’s hockey team needed a goal in the final minute to beat Pakistan on Thursday. Captain Harmanpreet Singh scored twice inside the first quarter and Sukhjeet Singh made it 3-0, before Pakistan clawed back level through second-half goals from Abu Mahmood, Hannan Shahid, and Waheed Rana. With a penalty shootout looming, Abhishek struck in the 60th minute to win it 4-3 and send India through to the gold medal match, where they will face the winner of Malaysia and South Korea.

A day earlier, the women’s team had already booked its own final. India beat South Korea 2-0 in the semifinal, Ishika converting a penalty corner in the 43rd minute before Salima Tete’s cross deflected in four minutes later. India will play defending champions China for gold on October 2.

It is the first time both Indian hockey teams have reached an Asian Games final in the same edition since 1998, when the men won gold against South Korea on penalties and the women finished runners-up to the same opponents. Both squads went through their group stages unbeaten this year, the women winning all five of their pool matches by a combined 66 goals to 3.

A final India hadn’t reached since 1962

Also on Thursday, wrestler Nitesh Siwach became the first Indian in 64 years to reach a Greco-Roman wrestling final at the Asian Games. He beat Mongolia’s Gankhuyag Ganbaatar 9-0 by technical superiority in the semifinal of the men’s 97kg category, India’s first appearance in a Greco-Roman final at the Games since 1962.

In the gold medal match, Siwach lost 0-4 to Japan’s Yuri Nakazato and had to settle for silver. It remains the best Asian Games result by an Indian Greco-Roman wrestler in more than six decades, in a style the country has historically entered far less often than freestyle wrestling.