GPT-6 Astra: OpenAI Declares "the AGI Era" With the Model It Paused a Month Earlier

On September 3, OpenAI released GPT-6 Astra, and company president Greg Brockman closed the presentation with "welcome to the AGI era." It's the same model we wrote about a month earlier: on August 7, OpenAI paused part of its work on Astra because internal testing couldn't rule out that the model could independently find and exploit zero-day vulnerabilities in hardened systems without human involvement. A month later, that same model hits the market under a banner about general intelligence, while its strongest offensive cyber capabilities stay locked away, just with a name now: Daybreak Access.

From pause to launch in four weeks

Astra was built on OpenAI's largest training run ever, using more than 100,000 GPUs at the Stargate complex in Texas. The company describes it as its most advanced model for complex work: coding, computer use, research and multistep tasks. Input and output token pricing sits at $10 and $50 per million, noticeably higher than the $4 and $20 for GPT-5.6 Sol, the company's previous strongest model.

On ExploitBench, a benchmark measuring the ability to independently exploit security vulnerabilities, Astra scores 100% against 78.5% for Sol. That's the exact capability that triggered the pause a month earlier. On OSWorld 2.0, a computer-use test, the score is 72.6% against Sol's 65.7%, at roughly 47% less time per task. On FrontierMath Tier 4, the hardest tier of research-level math problems, Astra posts 97.6% against Sol's 83%.

The number the whole "AGI era" rests on comes with fine print

The headline result OpenAI showed as evidence of crossing the AGI threshold was 99.9% on ARC-AGI-3, a benchmark measuring general reasoning on novel tasks. That score, though, comes from a test run on provider-specific infrastructure that preserves the model's reasoning state between actions. Under the standard, neutral test environment, without that advantage, Astra's score drops to 62.7%, at far higher compute cost for the same task. Thirty-seven percentage points of difference between the number from the press conference and the number under neutral conditions is exactly the kind of detail that disappears from headlines but decides whether "the AGI era" describes reality or describes marketing copy.

The benchmark's own creator won't back the claim

The sharpest pushback came from the organization behind ARC-AGI-3 itself. ARC Prize, the benchmark's creator, the one OpenAI built its AGI claim on, flatly declined to confirm that reading: nobody involved in building the test, including its own authors, is claiming Astra meets the definition of artificial general intelligence. Toby Walsh, chief scientist at the University of New South Wales' AI Institute, went further, saying he'd eat his own hat if Astra didn't trip up on trivial tasks an eight-year-old handles without effort. Rebecca Johnson of the University of Sydney added that OpenAI's own definitions of AGI expose the weaknesses behind this claim once you actually hold Astra up against them.

The launch also landed in the middle of an unusually packed week for the whole industry. On Tuesday, Anthropic showed off Fable 5.1 and Mythos 5.1. On Wednesday, Meta announced Muse Spark 1.3 and Google unveiled Gemini 3.8 Flash. Thursday's Astra closed out a run of four major launches from four different labs in five days. Trade press has already started calling this "model fatigue": the pace of updates is disorienting even companies buying access in bulk, and more than 1,100 AI lab employees separately signed a petition to Washington asking for frontier model development to slow down. In a week like that, "the AGI era" doubles as a way to cut through a flood of competing announcements.

Daybreak Access: the same logic as Anthropic's, under a different name

Astra's strongest cybersecurity capabilities remain available only through the Daybreak Access program, limited at launch to a narrow set of organizations. OpenAI says it plans to expand access gradually, eventually with lighter safeguards, covering vulnerability and proof-of-concept validation, malware analysis and detection engineering. Regular ChatGPT Plus, Pro, Business and Enterprise users were told they'd get access to Astra itself "in the coming days" after launch, not immediately.

That architecture, a public model plus a separate, controlled program gating the riskiest capabilities, is exactly the same solution Anthropic landed on two days earlier with Claude Fable 5.1 and Mythos 5.1. Two competing companies, 48 hours apart, published nearly identical answers to the same question: what to do with a model that can break security better than most humans. The difference is cosmetic: Anthropic built a formal verification program jointly with the US government, while OpenAI is keeping Daybreak Access in-house for now, with no disclosed institutional partner.

A launch that briefly contradicted itself

The launch moment itself was chaotic. OpenAI's launch materials hit the web before the model's official page even worked, so reporters at Reuters, CNBC and other outlets were quoting a post the company briefly pulled down in the middle of its own launch. The page came back a few dozen minutes later. It's a small incident, but it captures how blurry the word "launch" has become in AI: a model can exist in several states at once, some journalists get it under embargo, some companies get early access, and the official announcement disappears and reappears before anyone can fully track it.