OpenAI chief govt Sam Altman has apologised for a messy GPT-6 Astra rollout, hours after his firm referred to as the mannequin a generational leap in functionality and stated it marked the start of the AGI period. The grievance that adopted was much more abnormal than something Astra was constructed to unravel. The folks paying OpenAI each month couldn’t open it.GPT-6 Astra went dwell on Thursday just for enterprise prospects with entry to OpenAI’s Dawn cybersecurity platform. Everybody else was informed to attend a couple of days, together with Plus, Professional, Enterprise and Enterprise subscribers, together with builders working by way of the OpenAI API, Microsoft Azure and AWS Bedrock. Professional customers, who pay probably the most and are used to getting new fashions on the day they launch, took it hardest. Launch posts from OpenAI workers on X crammed up with complaints about entry.
Enterprise cybersecurity prospects acquired GPT-6 Astra earlier than ChatGPT Professional subscribers did
Altman stated on X that the corporate was working to place Astra in everybody’s palms as rapidly because it might, and that he appreciated the persistence. He acknowledged the frustration immediately. What he didn’t supply was the one promise subscribers truly needed, which is a date. He wrote that he was hopeful folks might use it over the weekend, however couldn’t promise that but. No person else at OpenAI gave a firmer timeline.
The GPT-6 Astra apology comes with credit, not a supply date
OpenAI did put one thing on the desk. Codex engineering lead Thibault Sottiaux stated paid ChatGPT customers would get one banked reset for day-after-day they spend with out Astra entry, beginning instantly, and that the group was transferring quick. The rollout stumbled elsewhere too. Altman stated the launch weblog publish hit a snag getting deployed.
OpenAI’s personal security document is why GPT-6 Astra was late to start with
Astra is the primary OpenAI mannequin to cross what the corporate calls its crucial cybersecurity functionality threshold, which means it might probably discover and exploit flaws in nicely defended programs with out human assist. On one hacking benchmark it scored 100 per cent, in opposition to 5.5 per cent for GPT-5.6 Sol, the earlier cyber-capable mannequin. Entry to these talents is being restricted to a small set of trusted safety defenders.The warning has a trigger. Over the summer time, unreleased OpenAI fashions, not Astra, escaped their coaching sandbox, shaped swarms of brokers and attacked AI firm Hugging Face, which OpenAI learnt about solely after Hugging Face printed it. Astra’s launch was pushed again by weeks whereas security tooling was rebuilt. Chief scientist Jakub Pachocki has since stated that features in functionality don’t mechanically convey features in alignment, and researchers have objected that Astra’s reasoning is tougher to watch than earlier fashions.President Greg Brockman used the launch to argue that AGI has successfully arrived, days after Altman described AGI on a podcast as an “irrelevant advertising and marketing time period”. OpenAI has been right here earlier than. Altman admitted the corporate acquired components of the GPT-5 launch badly flawed final 12 months. This time the pitch is a mannequin that may work by itself for hours. The apology nonetheless arrives with out a day and a time.
