OpenAI's President Just Said the Words "AGI Era" Out LoudGPT-6 Astra
Published: September 6, 2026 | Updated: September 6, 2026 | Category: AI | By Mahesh
OpenAI released GPT-6 Astra on September 3, and in a press briefing with reporters ahead of the launch, company co-founder and president Greg Brockman made a claim frontier AI labs have historically avoided stating so directly. "I think it's not unreasonable to feel that we are now in the AGI era," Brockman told reporters, according to Fox Business's account of the briefing.[1] He closed the session more plainly, according to multiple outlets present, including Axios: "Welcome to the AGI era."[2]
Brockman's framing, read closely across the outlets that covered the briefing directly, stops short of a formal declaration. According to The Next Web's reporting, which cited a Washington Post account of the same briefing, Brockman said "I do leave it up to the reader to decide for themselves if this qualifies for them," before adding, "I think we're there."[3] That is a deliberately hedged formulation: personally asserting a belief while explicitly declining to make an institutional claim on OpenAI's behalf, a distinction that matters given how consequential a formal AGI declaration would be, both technically and contractually, for a company with financial agreements tied to that exact term.
What OpenAI Says Astra Can Actually Do
Behind the AGI framing sits a specific, concrete product change that OpenAI is positioning as the real substance of this release: computer use. According to VentureBeat's detailed reporting from the briefing, Brockman described Astra as able to "zip through spreadsheets, fill out forms, and navigate across web pages often at superhuman speed."[4] OpenAI's own demonstration video, described by Axios, showed the model formatting a legal contract, building a 3D game, searching for food, and booking a tennis court, reportedly while handling multiple tasks in sequence rather than waiting for step-by-step human direction between each one.
The specific benchmark figures OpenAI disclosed give some texture to how much of a jump this represents from the company's own prior flagship. On an offline subset of OSWorld 2.0, a benchmark measuring an AI agent's ability to complete real computer tasks, Astra scored 72.6% while taking roughly 40 minutes per task, compared with predecessor GPT-5.6 Sol's 65.7% at roughly 75 minutes, according to VentureBeat's reporting, a roughly 47% reduction in time per task alongside a higher success rate. OpenAI also reported a 99.9% score on ARC-AGI-3, a benchmark specifically designed to test a model's ability to reason through situations it has not previously encountered, according to Fortune's reporting, which also noted OpenAI's benchmarks show Astra outperforming Anthropic's newly released Claude Fable 5.1 across a range of tasks.[5]
The Cybersecurity Story This Launch Was Already Tangled In
This release does not arrive as a clean, standalone story, and OpenAI's own framing acknowledges as much. Depth Grid covered in detail earlier this month how OpenAI confirmed Astra was the first model to officially cross the Critical cybersecurity threshold under the company's own Preparedness Framework, and how that finding triggered a two-week internal pause in training. This week's launch confirms how OpenAI ultimately resolved that safety finding for public release: according to The Next Web's reporting, both the version going to Daybreak cybersecurity program participants and the version reaching general ChatGPT subscribers block access to the model's most advanced cyber capabilities, with the consumer-facing version carrying additional guardrails specifically layered on top.
Sam Altman, in a separate interview reported by Fox Business, directly addressed a related security concern from the same period, telling Fox Business anchor Liz Claman that an earlier test version of the model had been connected to a security breach affecting Hugging Face's systems, a detail that OpenAI had previously and separately confirmed in its own safety disclosures. Altman used the interview to reiterate that the publicly released version of Astra carries strict security standards distinct from whatever configuration was involved in that earlier incident. Reading these two threads together, the cybersecurity capability disclosure and this week's product launch, clarifies that Astra's release was not a simple model upgrade announcement; it was the resolution of an internal safety review process that had already become public before the launch itself happened.
Why Brockman's Timing and Wording Both Matter
The AGI framing did not emerge in a vacuum, and the launch's timing relative to OpenAI's closest competitor is a detail worth noting explicitly. According to reporting citing NBC News' coverage of the launch, Astra arrived just two days after Anthropic released its own Claude Fable 5.1 and Mythos 5.1 models, which Anthropic had itself described as "the world's most advanced models for coding and knowledge work."[7] OpenAI's own announcement described Astra in almost identically sweeping terms, calling it "state-of-the-art on computer use, browser use, software engineering, cybersecurity, science, and professional work" and "the world's most intelligent and aligned model." Two frontier labs making maximal capability claims about their flagship models within 48 hours of each other is itself a signal of how competitively compressed the current AI release cycle has become.
VentureBeat's reporting captured a more candid moment from the briefing that complicates the confident public messaging. OpenAI vice president of research Mia Glaese, discussing the same computer-use capability Brockman was promoting, acknowledged that OpenAI has been working toward this specific goal for years, describing the launch as showing "how far we've come from sort of aspirationally training for computer use to bringing real value to people every day." That framing, real progress toward a long-standing goal rather than a sudden, discontinuous leap, sits in some tension with Brockman's own "generational leap" and "AGI era" language from the same briefing, and multiple outlets covering the launch noted that gap directly rather than taking the more dramatic framing at face value.
What the Voluntary Government Review Actually Confirmed
Before release, Astra went through what The Next Web's reporting described as a voluntary US government prerelease vetting process for advanced AI systems, established under an executive order issued in June 2026. Brockman told reporters the review process did not require any changes to the model's safeguards, a detail he offered as reassurance about the model's safety. The Next Web's own reporting flagged an important caveat to that framing, however: participation in the review is voluntary, the National Security Agency director decides which models even qualify as covered frontier models subject to review through a classified benchmarking process, and no formal approval mechanism exists for the government to grant or withhold in the first place. In other words, the absence of requested changes reflects the absence of a review process with the authority to demand them, not necessarily an independent government judgment that the model's existing safeguards were sufficient.
What to Watch Next
GPT-6 Astra's rollout is happening in stages, with Daybreak program cybersecurity customers receiving access first on launch day, followed by ChatGPT Plus, Pro, Business and Enterprise subscribers "in the coming days," and eventual availability through the OpenAI API and cloud platforms including AWS Bedrock and Microsoft Azure, according to BetaNews's reporting. Free-tier users and OpenAI's cheapest paid tier will not receive access in the near term. The clearest signal to watch over the coming weeks will be independent, third-party verification of the specific benchmark claims OpenAI disclosed at launch, since every performance figure discussed this week originates from OpenAI's own internal testing rather than an outside evaluator, a distinction that matters significantly for a launch built substantially around the claim that a genuine capability threshold has been crossed.
Read More on Depth Grid
- OpenAI built an AI that can hack anything on its own. Then it had to decide whether to ship it.
- Nvidia now owns the platform that decides which AI models the world actually uses
- Broadcom's AI chip revenue tripled in a year. Investors sold the stock anyway.
- OpenAI wanted $2.5 billion in ad revenue this year. It's on pace for $1 billion.
- Anthropic's IPO paperwork is about to admit, in writing, that people don't want its data centers
Article by Depth Grid News Desk | depthgrid.in