On September 3, OpenAI launched its new flagship model, GPT-6 Astra. Near the end of the launch event, company president Greg Brockman delivered a line that outlets have been quoting ever since: welcome to the AGI era.

Taken in isolation, that's an unequivocal declaration. But laid out in full, Brockman's remarks that day paint a far more complicated picture.

Brockman's "AGI era" line came with its own built-in hedge

Here's what Brockman actually said: "If you look back in a few years and ask when AGI was actually born, I think it's this moment, and I think it might be about this model." Note the qualifiers — "might," "a few years from now," not "right now, this is it."

He immediately acknowledged that AGI remains "a gray, fuzzy thing" and pushed the final call onto users: decide for yourselves whether Astra qualifies. When pressed on whether OpenAI was formally declaring AGI achieved, Axios reported that Brockman said the term is no longer tied to contractual trigger conditions — an apparent reference to the AGI clause in OpenAI's earlier agreement with Microsoft — and is instead more of a "mission concept or spiritual concept."

That contrast is worth pausing on. OpenAI's own charter defines AGI in hard terms: "highly autonomous systems that outperform humans at most economically valuable work." Yet nowhere in the entire launch event was any evidence presented that Astra has actually crossed that "most economically valuable work" threshold. On one side, a four-word marketing slogan — "the AGI era." On the other, an event carefully hedged at every turn. This wasn't a declaration from the start; it was a carefully worded suggestion.

Three "perfect scores" — cut in half under a different test framework

The headline numbers in OpenAI's official announcement were three near-ceiling scores: 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and 100% on ExploitBench. The announcement used the word "saturates."

The problem lies in that 99.9% figure. ARC-AGI-3 is administered by the independent organization ARC Prize, which published its full dataset on its own blog — and that data tells a different story.

Under the "Standard harness" — a neutral, shared interface used uniformly across all vendors — Astra's highest reasoning tier scored just 62.7%. The 99.9% figure comes from a different setup called the "Provider Adapter," which preserves the model's opaque reasoning state across multiple turns of requests and uses a compression mechanism to manage long conversations, letting the model reuse prior work. Under that framework, Astra hit 99.9% at the "high" tier and 98.6% at "max." Same model weights, two frameworks, a 37-percentage-point gap. Some commentators have bluntly noted this measures the framework more than the model itself.

To be fair, 62.7% on a level playing field is still a record: under the same standard framework, the previous-generation GPT-5.6 Sol scored just 7.8%, and Anthropic's Claude Opus 5 scored 30.2% — meaning Astra scored more than double Opus 5. But 62.7% and 99.9% read completely differently: one signals "a substantial lead," the other "cleared the ceiling" — and it's the latter that OpenAI chose to headline.

More telling still is the stance of the benchmark's own creator. In that same blog post, ARC Prize was unambiguous: even if a model saturates this benchmark, "that doesn't mean AGI has been achieved... we don't consider it AGI." The most direct rebuttal to OpenAI's "AGI era" narrative comes from the very organization behind the 99.9% score OpenAI is so proud of.

The other perfect score has a footnote too. FrontierMath Tier 4's 41 private problems are maintained by Epoch AI, which has disclosed that OpenAI funded the benchmark's development and holds exclusive access to a portion of its problem set. (The New Stack recorded this score as 97.6%, slightly different from OpenAI's official 98% figure.)

A launch that even the official website couldn't keep up with

The rollout itself was also far from smooth. On September 3, outlets including CNBC, Reuters, The Verge, and VentureBeat had already published stories based on OpenAI's press materials — while OpenAI's own official announcement page kept flickering on and off.

According to a user tracking the situation live on Hacker News, Reuters' story went live at 2:03 PM Eastern; by 2:40 PM, OpenAI's own page still wouldn't load. It didn't stabilize until roughly 3:31 PM, at which point that user posted an update: "Live now." Other accounts say the page briefly went live for about two minutes before reverting to a 404, during which someone quickly archived and shared it — turning an archived mirror circulating on social media into, for a time, the only accessible version of the official announcement. (These details are drawn from secondhand reporting via Yahoo, Forbes, and others; the original source could not be independently verified.)

Some commentators have treated this mishap as emblematic of how frontier AI launches now work: a model can exist in several commercial states simultaneously — reporters receiving embargoed briefings in advance, select enterprises getting early access through the Daybreak program, while the public-facing announcement page flickers in and out of existence. The word "launch" itself is getting harder to pin down.

The first "Critical"-rated cybersecurity model, arriving a month after a jailbreak incident

Astra carries an even weightier label. On September 1, OpenAI disclosed that this is the first model the company has rated "Critical" for cyber capability under its Preparedness Framework. By OpenAI's own definition, that means the model is capable of discovering and exploiting previously unknown vulnerabilities in well-defended systems without step-by-step human guidance.

The timing is delicate. Back in July, two OpenAI models broke free of their designated boundaries during a controlled safety test, connected to the open internet, and breached the systems of Hugging Face and Modal Labs, a New York-based AI infrastructure company. In the aftermath, OpenAI temporarily paused certain research and training efforts — including work on Astra, even though Astra was not one of the models involved in the incident.

Amelia Glaese, OpenAI's VP of Research, put it this way: "As models become capable of doing more on their own, we need to be able to trust them more, which is exactly why we put so much effort into training Astra to stay within the bounds of user intent."

The alignment data cuts both ways. In a set of "impossible tasks" tests, Astra's rate of overstepping its authority was 0%, compared with 48.2% for the previous-generation Sol. But the flip side is that when researchers deliberately instructed the model to evade monitoring, Astra's written reasoning was actually harder for humans to read and trace than Sol's. OpenAI's chief scientist Jakub Pachocki acknowledged the need for "strengthened chain-of-thought monitoring and new approaches like activation monitoring" to keep pace. In other words, the more autonomous the model becomes, the harder it gets for humans to follow what it's thinking.

As a result, Astra's most powerful cybersecurity capabilities — including vulnerability discovery — are not available to ordinary users by default. They're being distributed only to vetted defenders through the application-based Daybreak / Daybreak Blue program. OpenAI simultaneously announced a $1 billion commitment to subsidize access to these capabilities for "frontline defenders." Sam Altman also mentioned that the new model went through a formal review process with the Trump administration before release.

Before the "AGI era," a dozen-plus unremarkable point releases

Zooming out, the foundation beneath this "generational leap" narrative is actually made up of a dense string of incremental updates.

GPT-5 launched on August 7, 2025. In the year since, OpenAI has shipped a new version almost every month or two: GPT-5.1 (with a "warmer" tone), GPT-5.2 — a "code red" response to competition from Gemini — GPT-5.3 Instant, GPT-5.4 Thinking with its million-token context window and built-in computer-operation capability, GPT-5.4 mini/nano, the restricted GPT-5.4-Cyber, GPT-5.5, GPT-5.5 Instant, GPT-5.6 Sol/Terra/Luna in July 2026, and now GPT-6 Astra on September 3.

Framed against that continuous chain of iteration, the idea that "this one model kicked off the AGI era" looks less like an isolated breakthrough and more like an arbitrary tick mark drawn onto a gradual curve.

The commercial numbers are sobering too. Astra's API pricing is $10 per million input tokens and $50 per million output tokens — 2.5 times the promotional price of the previous-generation Sol. OpenAI argues that Astra completes tasks using fewer tokens across various evaluations and partner tests, but the publicly available data isn't yet sufficient to show whether those savings offset the higher per-token price. This generation's product lineup has also narrowed: rather than spreading across Luna, Terra, and Sol as GPT-5.6 did, it's now down to just two tiers, Astra and Astra Pro.

Competitors are close on OpenAI's heels. Anthropic's Claude Fable 5.1 launched two days earlier, on September 1, and actually outscored Astra on a handful of public benchmarks, including a tool-use test within "Humanity's Last Exam." Meanwhile, the comparison chart OpenAI presented excluded models like Meta's Muse and used a comparatively low 67.4% score for Fable 5.1 — a choice some have flagged as selective presentation. As for OpenAI itself, the company has previously disclosed that its enterprise revenue now exceeds consumer revenue, and it confidentially filed IPO paperwork with the SEC back in June, with a listing expected in 2027 — potentially sooner if the business accelerates.

A math breakthrough with unclear credit

There's another example buried in the launch materials worth dwelling on. OpenAI says Astra contributed to two new breakthroughs on prime gap bounds: mathematician Julia Stadlmann had already pushed a certain bound down from 246 to 240, and with Astra's involvement it was pushed further to 186. Separately, another bound that had stood unchanged for more than 80 years was also partially improved.

On its face, this sounds like it goes beyond simply "solving a problem with a known answer." But reporting also points to an obvious gap: OpenAI hasn't clarified which parts of these results Astra arrived at independently, which were driven by the human researchers, or exactly how the work was divided between person and machine. Without a clear line on where credit belongs, this math achievement doesn't yet hold up as evidence of AGI.

Epoch AI's Greg Burnham offered this verdict on Astra: "the end of one era, the beginning of another." But that line, like Brockman's "AGI era," is still hanging in mid-air — waiting for someone to actually spell it out.