OpenAI says GPT-6 Astra could mark the beginning of the AGI era. The more uncomfortable possibility is that it may also mark the beginning of an era in which AI becomes harder to understand.
There is something different about GPT-6 Astra. OpenAI is not presenting its latest model simply as a smarter chatbot: Astra is designed to use computers, navigate software, operate websites, work with spreadsheets and carry out complex tasks with far less human intervention. The company calls it the world’s best computer-use model.
It is also the first OpenAI model to reach what the company describes as a “Critical” cybersecurity capability threshold, and president Greg Brockman has suggested that, with hindsight, the launch could eventually be seen as the moment the industry entered the AGI era. That is an extraordinary set of claims.
Depending on which benchmark you choose, GPT-6 Astra looks either like a genuine generational leap or a more modest improvement over its predecessor. And as its reasoning becomes more sophisticated, researchers are increasingly concerned that some of what happens inside the model may become harder to observe.
AI that does not just answer — it acts
For years, the basic interaction with an AI model has been remarkably simple: you ask a question, and it gives you an answer. Astra is designed to go considerably further. It can interact with a computer in ways that resemble a human user, navigating websites, manipulating spreadsheets and working with desktop applications rather than relying on a separate API integration for every action.
That matters because the next stage of AI development is increasingly about agents. The goal is no longer simply to produce a good answer, but to give an AI system a task and allow it to work out the steps required to complete it.
OpenAI’s benchmark results suggest Astra is particularly strong in this area. The model scored 72.6% on OSWorld V2-Offline, compared with 65.7% for GPT-5.6 Sol, and 74.1% on DeepSWE v1.1 against around 71% for Sol. On Terminal-Bench 4.0 the gap was considerably larger: 57.9% against 37.3%. OpenAI also reported 97.6% on FrontierMath Tier 4 v2, 96% on GPQA Diamond, 95.9% on BenchCAD and 100% on ExploitBench.
Taken together, the numbers make for an impressive launch presentation. But benchmarks come with a rather important footnote.
The ruler matters
AI companies have become exceptionally good at demonstrating impressive benchmark improvements. The problem is that benchmarks are not neutral measures of intelligence.
Take FrontierMath. As Vellum has noted in its analysis of the launch, Epoch AI, which administers the benchmark, has recorded that OpenAI funded its development and has exclusive access to part of the evaluation. That does not make Astra’s result meaningless, but it does mean the result deserves to be considered in context when OpenAI uses it to demonstrate superiority.
There are other complications. OpenAI’s announcement reports a 99.9% score for Astra on ARC-AGI-3, while The New Stack reported 98.6%, apparently using a different testing configuration. And on DeepSWE, Meta’s Muse Spark 1.3 reportedly achieved 75.4% at maximum reasoning — around 1.3 percentage points higher than Astra. That is not a dramatic difference, amounting to roughly one or two tasks out of 113, but it is enough to complicate the claim that Astra is simply the best model in every meaningful sense.
The most interesting comparison comes from Artificial Analysis, whose independent Intelligence Index gave GPT-6 Astra a score of 61. That puts it level with GPT-5.6 Sol, while leaving it behind Meta’s Muse Spark 1.3 and Anthropic’s Claude Fable 5.1.
This is not necessarily evidence that OpenAI’s benchmark results are wrong. Rather, it demonstrates how much the answer depends on the question being asked. If you measure computer use, coding or specific reasoning tasks, Astra can look extraordinary. If you use a broader independent intelligence index, the generational leap becomes considerably harder to see.
The question is not only how well AI performs. It is who gets to decide what “well” means.
The most worrying capability may be the one involving hackers
If Astra’s benchmark results are impressive, its cybersecurity capabilities are potentially far more consequential. OpenAI says the model is the first to reach its Critical cybersecurity threshold, which in practical terms means Astra can identify previously unknown vulnerabilities and develop functional exploits without requiring a human to guide it through every step.
During testing, it reportedly discovered and chained two zero-day vulnerabilities, which OpenAI notified to the relevant maintainers. In isolated target tests, Astra compromised 10 out of 22 targets, where GPT-5.6 Sol managed just one. The difference was similarly striking in reverse engineering: asked to analyse binaries without source code, Astra solved 99.2% of the challenges within four attempts, compared with 68.7% for Sol.
This is more than a better coding assistant. A model capable of discovering vulnerabilities, understanding unfamiliar systems and constructing working exploits autonomously represents a significant shift in the cybersecurity landscape — with the obvious benefit being defensive research, and the obvious risk being that the same capability could be used offensively.
OpenAI says the public version of Astra has been trained to refuse advanced cybersecurity requests, including requests to write proof-of-concept exploits, while verified organisations participating in its Daybreak cybersecurity programme receive less restrictive safeguards for legitimate security work.
But the company is clearly expecting problems. Mia Glaese has warned that users should expect slowness, pauses and blocks at launch, potentially even during tasks that have nothing to do with cybersecurity. OpenAI says Astra’s model files are encrypted and actively monitored, and its API security system can interrupt a task rather than merely pausing it for review.
What happens when the reasoning disappears?
The cybersecurity story is worrying enough. But another development around Astra could have even greater implications for the future of AI: according to The Information, reported by TechCrunch, OpenAI is using a technique known as recurrent depth, sometimes described as opaque recurrence.
The idea is relatively straightforward. Instead of making a model reason for longer by generating more visible tokens, the system can repeatedly process information internally. That could make models more efficient and allow them to perform more complex reasoning without producing correspondingly longer chains of thought.
But there is a catch. For researchers studying AI safety, the chain of thought has become one of the most useful — if imperfect — windows into model behaviour, and if more reasoning happens inside internal representations that are never exposed as text, that window becomes smaller.
Buck Shlegeris of Redwood Research has expressed serious concern that scaling recurrent reasoning could undermine chain-of-thought monitoring. His colleague Ryan Greenblatt has suggested the natural progression could ultimately be AI systems carrying out most, or even all, of their reasoning in latent space. Zvi Mowshowitz has described the direction as “playing with fire”, warning of a potential race towards less transparency.
OpenAI disputes the idea that Astra represents an immediate disappearance of interpretable reasoning. Jakub Pachocki has said the use of recurrent processing in Astra is limited, that its chain of thought remains legible, and that the computational graph depth of current frontier models, including Astra, remains within roughly a factor of two of GPT-4.
That may be reassuring, but the underlying direction is difficult to ignore. The more sophisticated AI becomes, the more we may be asking it to reason in ways that humans cannot directly observe — and that creates a dangerous asymmetry, in which we become better at building intelligent systems faster than we become better at understanding them.
So, is this AGI?
Greg Brockman has suggested that GPT-6 Astra could eventually be remembered as a turning point — perhaps even the beginning of the AGI era. It is the sort of statement guaranteed to attract attention, and it immediately raises a deceptively simple question: what exactly is AGI?
There is still no universally accepted definition. OpenAI’s founding charter described AGI as systems capable of outperforming humans at most economically valuable work, yet Brockman has also argued that AGI is no longer a particularly useful concept in practical terms. That apparent contradiction is partly connected to OpenAI’s changing relationship with Microsoft: an earlier agreement gave Microsoft exclusive licensing rights until AGI was reached, and those terms were later scrapped.
So we are left with an unusual situation. OpenAI is willing to talk about Astra as a possible milestone on the road to AGI while simultaneously arguing that AGI itself is no longer a particularly useful threshold. Meanwhile, an independent evaluation puts Astra on the same Intelligence Index score as its predecessor. Perhaps that is the clearest illustration of the uncertainty surrounding the entire debate.
Astra may matter for reasons beyond intelligence
It is tempting to reduce GPT-6 Astra to a familiar question: is it smarter than the previous model? That may be the wrong question. The more important development is that Astra is moving AI further away from being something that simply responds and closer to something that acts — operating software, completing multi-step tasks, reasoning about unfamiliar environments, finding vulnerabilities, and increasingly doing so without a human specifying every step.
But perhaps the most important lesson from Astra is not about intelligence at all. It is about visibility. As AI systems become more autonomous, their internal processes may become less accessible to the humans responsible for monitoring them. That creates a race of its own: not simply a race to build the most capable model, but a race to develop systems capable of understanding, auditing and controlling increasingly capable AI.
Are we building AI systems capable of doing more and more things before we have equally powerful tools for understanding everything happening inside them? If the answer is yes, then the most important legacy of GPT-6 Astra may not be its benchmark scores. It may be the moment when the gap between what AI can do and what we can see it doing became impossible to ignore.