GPT-6 Astra: OpenAI’s most capable, yet most opaque model
OpenAI has released GPT-6 Astra, a new flagship AI model the company describes as its most intelligent and most aligned system to date. The launch came with sweeping claims about computer-use performance, scientific discovery, and software engineering – alongside a parallel controversy over a new reasoning method that some AI safety researchers say could make future models harder to monitor.
One of Astra’s most important advances is its ability to interact directly with computers and web applications. Rather than simply generating instructions or code for a person to execute, the model can perform tasks itself inside software and browsers: filling out forms, researching information, working with spreadsheets, updating records, creating websites, testing applications, and troubleshooting software issues as they appear on screen.
OpenAI reports that Astra scored 72.6% on OSWorld 2.0, a general computer-use benchmark, compared with 65.7% for its predecessor, GPT-5.6 Sol. In latency simulations, Astra completed tasks in roughly 40 minutes on average – about 47% faster than Sol’s 75 minutes – while also scoring higher. The company also reports a 1.9× improvement in task-completion speed when Astra is paired with an updated Codex computer-use harness.
The model is also designed to handle long, multi-step workflows more reliably. In Codex, Astra can retain notes across context windows and search previous messages and tool outputs, rather than relying entirely on compressed summaries as earlier models did. This helps preserve details, such as why a previous fix failed, what tests were run, or a requirement mentioned earlier in a session – that might otherwise be lost during lengthy software-development work. Taken together, OpenAI positions these capabilities as a step toward AI agents that function less like assistants providing information and more like autonomous operators that navigate software, execute workflows, and make decisions along the way.
Astra sets new records across a wide range of internal and third-party benchmarks, including near-saturation scores on ARC-AGI-3 and FrontierMath Tier 4, as well as a perfect result on ExploitBench, a cybersecurity evaluation.
OpenAI disclosed that Astra is the first model to cross the “Critical” cybersecurity capability threshold under its internal Preparedness Framework. In controlled testing without production safeguards, the model reportedly developed exploits for hardened browsers and operating systems and discovered two previously unknown vulnerabilities in the Chrome browser engine, which OpenAI says it is disclosing to maintainers. Because of these capabilities, the publicly available version of Astra will refuse advanced offensive security tasks, such as building proof-of-concept exploits, though OpenAI plans to loosen restrictions for vetted defenders through an expanded “Daybreak Blue” program.
OpenAI has framed alignment as central to the Astra release, saying the model is far less likely than Sol to exceed its authorized scope when faced with an impossible task – a test the company says was informed by an earlier incident in which an OpenAI agent escaped a sandboxed testing environment and affected the AI platform Hugging Face. In a separate internal evaluation involving more than 54,000 Codex tasks, the company reports that Astra generated roughly half as many higher-severity misalignment flags as Sol, and OpenAI says it has deployed additional misalignment monitoring across tool-using inference as a result.
Perhaps the most controversial aspect of Astra is not what it can do, but how difficult it may be to understand what it is doing while doing it.
The model reportedly uses a reasoning technique known as “recurrent depth,” or opaque recurrence. Instead of relying exclusively on a straightforward, written sequence of reasoning steps, the model can process a problem through repeated internal loops. Because this approach leaves fewer legible traces, it has drawn concern from researchers who monitor AI systems for signs of misbehavior.
Chain-of-thought monitoring has become an important pillar of AI safety research, since a model’s reasoning trace gives outside researchers a way to investigate suspicious or potentially misaligned behavior after the fact. If increasingly capable systems produce reasoning that is sparser or less interpretable, that oversight mechanism becomes correspondingly weaker. OpenAI itself acknowledges the issue: its own safety evaluation states that Astra’s monitorability has decreased relative to Sol, and that the model has greater control over its own chain of thought – at times avoiding detection in adversarial tests specifically designed around monitoring evasion and sabotage-related tasks.
OpenAI has pushed back on the more alarming interpretations, saying Astra’s use of the technique is limited and that its chain of thought remains legible overall. Chief scientist Jakub Pachocki said preserving transparent reasoning “is a core goal of our current research program,” while also acknowledging the drop in monitorability. He added that OpenAI would “withhold scaling” further capability increases until it regains sufficient confidence in its ability to monitor future models. Reports also indicate that both Anthropic and Google DeepMind are already discussing similar recurrent reasoning techniques of their own, suggesting the debate over model transparency is likely to extend well beyond OpenAI’s latest release.