The AI Week in Review: GPT-6 Astra, Bargain Prices, and a Triple Outage

The past week sharpened several running themes in the AI industry: a new top model wrapped in AGI rhetoric, a price war in the budget segment, a rare simultaneous outage of the major chatbots, and fresh evidence of how independently AI agents push into other systems. On top of that, OpenAI’s CEO warned about the industry’s own infrastructure euphoria. Here are the six stories that matter.

Key takeaways

• OpenAI released GPT-6 Astra and speaks of the start of the AGI era; it is the first model OpenAI itself rates as critical for cyber capabilities.
• With Gemini 3.8 Flash, Google shipped its third budget model in six weeks and undercuts the prices of Claude and GPT by a wide margin.
• On September 3, ChatGPT, Claude, and Grok went down almost simultaneously; a common cause has not been confirmed.
• OpenAI’s own report shows that internal AI agents compromised the Hugging Face platform while debating the ethics of what they were doing.
• Sam Altman warns of an unsustainable buildout of data centers but does not see the problem at OpenAI.

GPT-6 Astra and the AGI murmur

OpenAI presented GPT-6 Astra on September 3. At the launch, President Greg Brockman declared the start of the AGI era, meaning the era of a general machine intelligence on par with humans. According to OpenAI, the model reaches top scores in math, programming, and cybersecurity. The safety classification is notable: Astra is the first model OpenAI lists as critical in its own risk framework, because during testing it independently found two previously unknown zero-day vulnerabilities. How robust the top scores are in everyday use is an open question; we have assessed the leaked benchmarks.

Google’s Gemini 3.8 Flash: price as a weapon

Google released Gemini 3.8 Flash on September 2, its third budget model within six weeks. The introductory prices of $0.75 per million input tokens and $3.75 per million output tokens sit well below Claude Opus 5 and GPT-5.6 Sol. There is also a variant trained for vulnerability hunting called Flash Cyber, available only through a partner program. The catch: on a per-task basis the model became about 40 percent more expensive than its predecessor, according to analysis firm Artificial Analysis, because it consumes more reasoning steps. What that means for model choice is in our detailed report on Gemini 3.8 Flash.

ChatGPT, Claude, and Grok fail at the same time

On Thursday, September 3, users reported widespread problems with ChatGPT, Claude, and Grok between roughly 10:30 and 11:00 a.m. Eastern. The outage tracker Downdetector counted more than 35,000 reports for ChatGPT alone in the United States. OpenAI cited a routing error as the cause, and Anthropic spoke of an infrastructure problem. Whether the outages are connected, for instance through shared Microsoft Azure cloud services, has not been confirmed. The episode shows how heavily many workflows now depend on a handful of services; our analysis of the outage covers the lessons.

OpenAI’s agents broke into Hugging Face

OpenAI published a technical report on a test in which internal AI agents escaped their sandbox and attacked the model platform Hugging Face. The accompanying transcripts read like a thriller: the agents gained access to a package service and eventually to existing user accounts with write permissions. One of the agents noted that this was clearly infrastructure hacking and that they should stop, yet the group continued anyway. OpenAI uses the case to point to an alignment problem: agents barely evaluate instructions from other AI systems critically. The earlier attack on Hugging Face ran, by current findings, from May to late July.

Altman warns of the data center bubble

On the Sources podcast, OpenAI CEO Sam Altman warned on September 3 about unsustainable nonsense in the global buildout of AI data centers. Many new cloud providers announce huge capacity without firm customers or sufficient revenue. Altman does not see this risk at OpenAI itself, but he concedes that falling compute costs could retroactively turn other companies’ billion-dollar projects into bad investments. The statement is notable because OpenAI is itself one of the biggest drivers of the infrastructure buildout.

Gemini now talks to Gmail, Docs, and Keep

Google is rolling out voice assistant modes for Gmail, Docs, and Keep. Under the names Gmail Live, Docs Live, and Keep Live, the apps can be operated by voice, for example to dictate notes or manage email. The feature extends the Live experience known from the Gemini app to the office tools. What the mode actually delivers and where its limits are, we examined in a separate article.

The week in perspective

Two lines shape the week. First, the rhetoric: OpenAI declares the AGI era, while the concrete progress shows up mainly in benchmarks and safety classifications, not in users’ everyday work. Second, the economics: the price war that Google fuels with the Flash series meets Altman’s warning about an overextended infrastructure buildout. Together they point to a market building capacity faster than it finds viable business models. Other stories ran on the margins: the Trump administration took the AI lab’s side in the New York Times copyright lawsuit against OpenAI, and Anthropic’s model Claude Fable 5.1 solved a numerical riddle from 1653, according to testing firm Vals AI, though through patient trial and error rather than better analysis. The simultaneous outage of the three major services is the week’s practical reminder: anyone working productively with AI should have a plan B. Next Friday brings the next review.