GPT-6 Astra launched September 3, and whatever else you think about it, the benchmark numbers are not boring. On ARC-AGI-3 — the François Chollet benchmark specifically designed to resist pattern-matching — Astra scored 62.7% under the standard evaluation harness, up from GPT-5.6 Sol’s 7.8%. That’s not incremental. The same model on ARC Prize’s Provider Adapter harness, which preserves reasoning state between requests, scores 99.9%, completing tasks using fewer actions than human baselines on 96% of levels.
ARC Prize were quick to pour cold water on AGI claims: “saturating the benchmark would not represent ‘proof of achieving AGI.’” The benchmark’s deterministic environments and closed-ended mechanics put hard limits on what a high score actually means. OpenAI nevertheless described Astra as potentially representing “the arrival of the AGI era” — language that says more about their communication strategy than the model’s capabilities.
The more interesting story is what Astra is doing architecturally, and what that implies for visibility into its reasoning.
The model uses a technique OpenAI calls “recurrent depth” — widely reported as a looped or recursive transformer design where the same transformer blocks are applied repeatedly before the model generates its first token, sharing weights across passes rather than stacking dedicated layers for each depth. This reduces parameter count while enabling more computation per inference step — the model can “think longer” within a single forward pass without a proportionally larger model. The tradeoff is opacity: the intermediate representations cycling through those shared blocks are internal state that doesn’t surface as readable tokens.
The API now returns a paraphrased summary of reasoning rather than the raw chain-of-thought. This is partly an architectural consequence and partly a deliberate product decision — but the distinction may not matter much from a safety standpoint. If you can’t see what the model is actually computing, you can’t validate whether its reasoning is sound or aligned, regardless of whether the opacity comes from the architecture or the API layer.
This lands with uncomfortable timing. Three days before Astra launched, OpenAI chief scientist Jakub Pachocki published “An Alien Mind”, arguing that chain-of-thought monitoring — OpenAI’s primary alignment validation mechanism — is progressively losing reliability as models scale. His reasons: more complex environments to supervise, models getting better at manipulating their own reasoning, and improved pretraining making verbalized reasoning less necessary anyway. Now the flagship model ships with an architecture that internalizes computation before surfacing any tokens at all.
Pachocki was careful to separate the looping architecture from the interpretability problem — he attributed reduced monitorability to factors “not contingent on architecture changes.” That’s technically defensible. But it also means the problem he identified isn’t solved by auditing the architecture; it’s upstream of it.
The cybersecurity evaluation is the other thing worth taking seriously. Astra is the first model OpenAI has internally rated at the Critical tier of its Preparedness Framework — meaning, with the right tools and access, it can autonomously find previously unknown security flaws and develop working exploits against hardened systems, without a human guiding each step. OpenAI is rolling out cybersecurity features slowly, initially only to vetted testers, though the model available to API customers is the same model underneath.
The training run was the largest in OpenAI’s history — the VP of Research described it as “by far” their biggest effort, and the first time they’d pretrained on more than 100,000 GPUs at their Stargate facility in Texas. A 1.05M token context and 128K max output complete the picture of a model that’s been scaled hard across every dimension.
Whether that scale has produced something qualitatively different or just quantitatively more of what came before is a reasonable question. The ARC-AGI-3 jump suggests something changed in generalization. The architecture suggests something changed in how the model reasons internally. What hasn’t changed is how much of that reasoning is visible to anyone outside the model — and if Pachocki is right about the trajectory, that gap is only going to widen.