OpenAI releases GPT-6 Astra: excitement over an "AGI era" and safety concerns

A new OpenAI model

OpenAI unveiled its latest AI model, GPT-6 Astra, on Thursday, September 3, calling it the "world's most intelligent and aligned model." According to Le Figaro, the model officially launched on Friday, September 4, while a BFM broadcast about its rollout was still discussing distribution on September 7 — suggesting a staged release, with priority access for companies enrolled in the "Daybreak" cybersecurity program.

Unlike earlier models that required users to issue instructions step by step, Astra can independently take control of a computer and complete tasks without constant human input — browsing the internet, filling out tax forms, booking appointments, building and testing a website, or developing a video game, Le Figaro reports.

"AGI has arrived," says Nvidia's chief

On Sunday, Nvidia CEO Jensen Huang declared on X that "AGI has arrived," crediting OpenAI's team and noting that Astra was trained on Nvidia chips. OpenAI president Greg Brockman said: "Welcome to the AGI era." Yet OpenAI's own CEO, Sam Altman, in an interview with the "Sources" podcast, called AGI "a very poorly defined term... almost an irrelevant marketing term," according to Business Insider.

Critics, including researcher Gary Marcus, argue Huang offered neither evidence nor a definition, and that by his own 10-point AGI criteria, Astra meets only one or two. Researchers behind the ARC Prize benchmark, which OpenAI cited when presenting Astra, also stopped short of calling it AGI, noting that the results came from tightly bounded test environments.

An impressive feat, and skepticism

According to Developpez.com, Astra completed Valve's puzzle game Portal entirely on its own — without human help or hints, relying only on screenshots — using 3,336 tool calls at a cost of $571.18. The playthrough officially took about two hours, but including the model's "thinking" time, the process stretched over nearly a full day. Skeptics note that Portal, released in 2007, has been extensively documented online, so complete walkthroughs were likely present in the model's training data.

Meanwhile, some AI-safety researchers are less concerned with what Astra can do than with how it does it: its internal "recurring reasoning" architecture produces fewer readable traces for safety monitoring before it acts — described by some voices cited by Le Figaro as "the worst development in AI safety yet." On BFM, experts also discussed an initiative by OpenAI scientists to slow the pace of AI development. The rollout itself was not smooth — Altman has already apologized for a chaotic launch day, during which some paying users got stuck on loading screens.