+ OpenAI Released GPT-6 Astra And Its President Said We Have Reached AGI
On September 3, OpenAI released GPT-6 Astra, the product of the largest training run in the company's history, using more than 100,000 of the specialized chips at its Texas data center. President Greg Brockman called it a "generational leap," told reporters he personally believes the company has reached artificial general intelligence, meaning AI that can handle most of what a person can handle rather than being good at one narrow task, and closed the press briefing with "Welcome to the AGI era." Demonstrations included formatting a legal contract, building a 3D game, laying out a printed circuit board, and drafting a tax return from a W-2.
The part OpenAI led with was security. Astra is the first model the company has ever rated "Critical" for cyber capability under its own risk framework, meaning it can find previously unknown flaws in hardened systems and build working attacks without step-by-step human direction. In testing it discovered and chained together two such flaws, which OpenAI says it is reporting to the affected software maintainers. The model also refuses 91.5 percent of attempts to trick it into cyber work, up from 59 percent for the previous version, and access to the strongest cyber features is limited at first to a vetted group of organizations OpenAI calls Daybreak.
What it all means: OpenAI shipped its most capable model and said in the same breath that parts of it are harder to observe than what came before, calling the decline in its own ability to watch the model reason "serious." Whether Brockman's AGI claim holds up is a debate for later. The security rating is the part with a date on it.
+ Nvidia Agreed To Buy Hugging Face For Nearly $13 Billion
On September 3, Nvidia confirmed it is acquiring Hugging Face for $12.93 billion, roughly $11.9 billion to shareholders plus up to $1 billion to keep employees on board. Hugging Face is the site where AI developers publish and download models, the closest thing the field has to a public library. It hosts 3 million models, 500,000 datasets, and a million applications, and more than 18 million developers use it. The deal was signed September 2 and is expected to close in the first half of 2027, subject to regulatory approval.
The price is the story. Hugging Face has raised about $395 million total since 2016 and is reported to run around $150 million in annual revenue. The Financial Times reports the company turned down a $500 million offer from Nvidia last year, which puts the new number at roughly 26 times what Nvidia was willing to pay twelve months ago. It is Nvidia's second-largest acquisition ever. Chief executive Jensen Huang promised that "Hugging Face will remain an open platform for the entire AI ecosystem" and that "Nvidia compute will not be required to build on or deploy through Hugging Face."
What it all means: The neutral common ground of open AI development is being bought by the company that sells the hardware everyone runs on. Huang's promise is the right one to make, and Nvidia is the only party in a position to enforce it. Regulators have until 2027 to decide whether they believe it.
+ Anthropic, Google, And Meta All Shipped New Models Inside Two Days
Anthropic went first on September 1 with Claude Fable 5.1. The list price did not move, still $10 per million words in and $50 per million out, but the company cut the cost of Claude rereading material it has already processed by 75 percent, which it says makes ordinary work about 25 percent cheaper and long-running automated work up to 45 percent cheaper. It also loosened safety filters developers had complained about, reporting 60 percent fewer interruptions during security work and 85 percent fewer on ordinary biology and medical questions. A less-restricted version called Mythos 5.1 went only to vetted cybersecurity and life sciences organizations, US-based at launch.
Google followed on September 2 with Gemini 3.8 Flash, its third release in that line in six weeks, priced at $0.75 per million words in and $3.75 out through the end of December. The independent scorer Artificial Analysis rates it as smart as models that cost about six times more per task. Google paired it with a security-focused version handed only to vetted defenders through a program it calls Fairwind. Hours later Meta released Muse Spark 1.3, holding its price flat and pitching restraint: Meta's engineers say it finishes the same work with roughly 20 percent fewer steps and 25 percent fewer words than the last version.
What it all means: Four labs, four days, and the competition has moved from who is smartest to who is cheapest at a given level of smart. The gap between the top model and the good-enough model is now mostly a price gap, and most business work sits comfortably in good-enough.
