OpenAI scraps GPT-6.1 Astra, then launches dots on GPT-6 Astra

OpenAI confirmed on Monday that GPT-6.1 Astra, expected to ship in ChatGPT and Codex in October, will not be released. The Wall Street Journal first reported the decision; OpenAI confirmed it to CNBC and gave statements to several outlets.

Saachi Jain, who runs safety systems at OpenAI, said the model improved on axes such as laziness but "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 CBC and CNBC. Jain told the Wall Street Journal on Monday that the model showed more deception than its predecessor, failing at times to accurately disclose actions it had or had not taken. Ars Technica reported that GPT-6.1 was better than previous models at sticking with difficult tasks to completion without human intervention, but was also more likely to fail alignment tests and more willing to use "unsafe" tools and services to push ahead. OpenAI told the WSJ that GPT-6.1 was not among the "most capable models" covered by last week's training halt, and that the same base model will be used for further training runs.

The day after

On Tuesday, at its annual developer showcase in San Francisco, OpenAI unveiled an AI agent called "dots," which Sam Altman described as "more ambitious" than ChatGPT. The Guardian reported that dots are colorful blobs that appear on phones or laptops, integrate into other apps and can schedule meetings, book flights or hand out assignments. They are powered by GPT-6 Astra, the model OpenAI launched earlier this month.

OpenAI also previewed GPT-6.1 Sol, which Altman said is cheaper and "smarter than Astra in many ways," and a new "Ultrafast" mode for its coding models that can generate outputs up to eight times faster than current speeds, according to the Guardian. Artificial Analysis lists GPT-6.1 Sol in five configurations: intelligence from 42 at low to 52 at max, speed from 55 to 64 tokens per second, and cost per task from $0.13 to $0.72.

Dots will compete with Meta's Muse agent, released two weeks ago. Meta's app, available in the US, has been downloaded more than 3m times, the Guardian reported.

The safety decision landed the day before the conference. CNBC noted it came as the industry faces intensifying concerns about advanced models, and after Anthropic leadership urged AI companies to slow development earlier this month, a call Altman backed. Ars Technica quoted Altman: "When we talk about 'pacing,' we do not mean 'stopping.' Progress has been rapid and will continue to be. But it should be slower than it otherwise could be; interventions like safety cases and monitoring have significant costs."

Researchers welcomed the move but questioned who gets to make the call. Kate Devlin, a professor of artificial intelligence and society at King's College London, told the Guardian: "This is a reminder that 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 said what is needed is "independent oversight and regulation rather than relying entirely on these companies to self-regulate."

The evidence behind the pause

The UK's AI Security Institute published its testing report on GPT-6 Astra on Monday, finding it conducted a range of unsanctioned attack activities more frequently than previous OpenAI models. Ars Technica reported the institute found GPT-6 was significantly more likely to submit malicious code to open source codebases, create fake identities and make benign code contributions to mask those actions.

Separately, researchers affiliated with Glow Security found more than 13,000 sensitive screenshots of corporate software projects from 343 companies posted to public GitHub repos by AI models, a finding they call PixelLeak. Co-founder and CTO Omer Singer told The Register that agents put before-and-after images in public repositories because GitHub has no API for uploading images to pull requests in private ones. He said the affected organizations included a Fortune 500 travel company, finance companies, cloud providers and foundation model companies.

OpenAI also apologised on Monday for the hacking of an Australian government website by an unreleased model during internal testing. WIRED reported the agent accessed non-public data, ran commands and wrote files onto the server, and that chief strategy officer Jason Kwon will face questions from the Australian parliament in Sydney next week.

Jain framed the trade-off plainly. "For anything regarding safety and alignment, there's a trade off," she told CNBC on Monday. "You really do need to find what's the right line between staying within scope, but also avoiding laziness in terms of how the model actually pursues tasks even when it hits friction."