OpenAI Calls Astra The Start Of The AGI Era. Its Own CEO Calls "AGI" A Marketing Term.

Greg Brockman closed OpenAI's launch briefing for its newest model with four words: "Welcome to the AGI era."

Days earlier, on a different podcast, OpenAI's own CEO had called the term AGI "at best a very poorly defined term," adding he thought it was close to an irrelevant marketing phrase. That's the tension running through GPT-6 Astra's entire launch: extraordinary claims, from the company's own president, about a term the company's own CEO doesn't think means much.

What Astra Actually Is

The Benchmark Number That Didn't Survive Contact With Independent Testing

OpenAI's headline claim was a 99.9% result on ARC-AGI-3, the benchmark built specifically to resist the kind of memorization that inflates AI test scores. It's the number that anchored the "generational leap" framing.

The ARC Prize Foundation, which built and owns the benchmark, ran its own evaluation using a standard test harness rather than OpenAI's own. Its result: 62.7%, roughly 37 percentage points below OpenAI's figure. The gap traces back to the harness, the surrounding software scaffolding that shapes how a model is allowed to attempt each task, not the underlying model weights.

• Analysts reading the gap say it points to Astra being meaningfully more token-efficient, not necessarily smarter at reasoning, than its predecessor

• That efficiency still matters commercially, since inference costs have become a real constraint on how much AI companies can deploy profitably

• For everyday coding work outside the benchmark, several reviewers have said Anthropic's Claude Fable 5.1 still performs more reliably

What The Launch Demos Actually Showed

OpenAI's own promotional material for Astra leaned heavily on a handful of set-piece demonstrations: the model extracting numbers from a W-2 and completing a federal tax return in a browser, laying out a circuit board, drafting a legal contract, building a 3D game, and running an apartment search through mapping tools.

The tax demo drew the sharpest pushback. Reviewers who examined it closely noted that the form shown on screen didn't match a real IRS 1040, and that the tax math itself came out wrong. OpenAI's own framing has consistently called this kind of output a draft meant for human review, not a completed filing, which is the more defensible claim, but it's also not what the demo visually promised at a glance.

Astra's First-Ever Rating: "Critical"

Separately from the benchmark dispute, Astra crossed a threshold no OpenAI model had reached before. OpenAI's Preparedness Framework ranks cyber capability across four tiers, low, medium, high, and critical, where critical means a model can find and exploit previously unknown vulnerabilities in hardened systems without a person guiding each step.

Astra is the first OpenAI model classified Critical. OpenAI says it delayed the model's release specifically to add safety testing once it determined Astra's cyber capabilities had crossed into that tier, and is limiting the most advanced version of that capability to vetted organizations in its Daybreak program. Independent security researchers have also pointed out, and OpenAI has disputed, that Astra closely resembles the model believed to be involved in July's Hugging Face breach.

The Reasoning Technique Worrying Safety Experts

The more technical concern has nothing to do with benchmarks. It's about how part of Astra reasons internally. According to reporting later confirmed in substance by OpenAI's own chief scientist, Astra uses a technique called recurrent depth, also called opaque recurrence or looped transformers, where part of the model loops the same internal layers over a hidden state multiple times before producing output, reasoning that never appears as text a human can read.

• Redwood Research CEO Buck Shlegeris said he was extremely concerned by the reporting, uncertain whether Astra is meaningfully less monitorable as a result

• Redwood researcher Ryan Greenblatt warned that scaling the approach further could eventually mean a model reasons entirely outside any readable trail

• Peter Wildeford of the AI Policy Network called the shift potentially reckless, noting that reading chains of thought was one of the only tools that let investigators reconstruct what happened during the July Hugging Face breach

OpenAI's chief scientist, Jakub Pachocki, said the technique is used in a limited, constrained way and that the company remains committed to chain-of-thought monitoring. He also acknowledged, in the same briefing, that the monitoring built to contain Astra's Critical-tier cyber capabilities is fragile and trending in a negative direction, and that a roughly 20% compute-overhead safety system will interrupt some legitimate work as a consequence.

Nobody Actually Agrees On What "AGI" Means

Brockman's framing assumed a shared definition that doesn't exist. Every major lab defines the term differently, and the differences aren't just academic:

That $100 billion profit metric is not a rhetorical detail. Under OpenAI's agreement with Microsoft, hitting a defined AGI threshold changes the terms of the intellectual property rights OpenAI has ceded to its largest backer. Whether a model qualifies as AGI isn't just a philosophical debate at OpenAI, it's a contractual trigger with direct financial consequences for the company and its executives.

Why This Landed Now

Astra's launch arrived in the middle of a compressed IPO race. Anthropic is reportedly targeting a valuation near $2 trillion as soon as October 2026, and has already had its own moment of market-moving attention around its Mythos model earlier in the year. Several commentators have read Astra's launch, and its AGI framing in particular, as an attempt to recapture that kind of attention before Anthropic's listing, rather than as a claim that stands entirely on its own technical merits.

None of that makes the underlying safety questions less real. Astra's Critical-tier cybersecurity rating and its use of harder-to-monitor reasoning would matter regardless of what's happening in the IPO market. But the timing is a reasonable thing to notice when evaluating how much weight to put on the loudest claims made at launch.

OpenAI's Astra: FAQ

Jans Bock-Schroeder

Publisher & Founder of AI Angst

Coming from the world of art, photography, and the luxury market, Jans launched AI Angst in 2025 to explore the cultural, ethical, and psychological impacts of artificial intelligence. His work bridges creative vision with critical technology analysis, offering clarity in an era of rapid technological change.