On September 3, 2026, OpenAI unveiled GPT-6 Astra, an update the company calls “the world’s most intelligent and aligned model.” OpenAI describes it as a generational leap forward, with capabilities that could bring AI closer to what some consider artificial general intelligence (AGI). But it also appeared to be one of the most controversial so far: the model is more autonomous, more capable of carrying out complex work, and powerful enough to raise serious questions about cybersecurity and how much control humans should retain.
OpenAI President Greg Brockman described it as a “generational leap forward”. He personally believes OpenAI could have achieved AGI with the launch of Astra, but left the final definition up to the community, without insisting on a clear label from the company. Users can assess for themselves whether the new model meets this status.
The new model is being rolled out in phases. First, it became available to a limited group of organizations, including companies participating in OpenAI’s Daybreak cybersecurity program. This week, the company plans to open the model to ChatGPT Plus, Pro, Business, and Enterprise users. It will also be available through the OpenAI API, Microsoft Azure, and AWS Bedrock.
What does the new model do?
OpenAI presents GPT-6 Astra as a new model for computing, programming, science, cybersecurity, and professional tasks. The key difference between Astra and its predecessors is that the model is not intended as a better chatbot, it is intended as a computer operator, focusing not on consulting, but on independent performance of complex professional work. If the chatbot era was about answers, Astra is about action.
The company says that one of Astra's main applications is working directly with a computer. The model can fill out online forms, update customer records in CRM, and organize a calendar. Astra can also analyze scientific data, build graphs, create websites, and run frontend QA. The model can install and test software on its own, as well as help solve problems it sees on the user’s screen. It can conduct research on the Internet, draft documents, and prepare summaries directly in email or a text editor.
It may all sound boring until you remember that most work is boring. It's often about opening five tabs, fixing a document, updating a CRM, comparing dashboards, checking a spreadsheet, making a presentation look decent, and writing a follow-up email. None of these tasks is particularly difficult on its own. The challenge is moving between different tools, keeping track of context, making decisions along the way, and checking that everything was done correctly. That's where Astra's ability to handle multi-step workflows becomes interesting. It can check its own work and handle long-running tasks, while also reasoning through science and math. And somehow, Astra can now even model houses in Blender and turn them into walkable Unreal Engine 5 scenes.
OpenAI demonstrated the model’s capabilities using examples from game development, electrical engineering, and everyday intelligent work. In the demo materials, the model formats a legal contract, creates a 3D game, and interacts with engineering and design environments - KiCad, Unity, FreeCAD, and Blender. Astra can also prepare a draft tax return based on W-2 data, effectively taking on routine bureaucratic tasks.
In one example, Astra performed PCB layout in KiCad. The model converted an electronic circuit into a PCB suitable for production, independently placing components and laying copper connections. The company notes that PCB layout is still often a manual task and can slow down the development of electronics. Automating this process lets engineers move faster to create, optimize, and test new designs.
How much better is Astra?
According to OpenAI, Astra shows results that outperform the previous GPT-5.6 Sol model and competitors in the market, including Anthropic’s Claude Fable 5.1 model. The company says the model combines years of research. It shows advanced results in computer and browser usage, software development, cybersecurity, scientific tasks, and professional work. They claim the model is better at navigating within the task at hand, understanding user intent, adhering to set constraints, performing long-running tasks, and working with multi-step workflows.
Astra scored 72.6% on OSWorld 2.0 at roughly 40 minutes per task, compared with GPT-5.6 Sol at 65.7% and roughly 75 minutes per task - that is, almost twice as fast with higher quality. So the claim is not just that it is smarter. The claim is that it gets more done, faster. That is the number businesses will care about.
Astra scored 97.6% on the FrontierMath Tier 4 benchmark, which tests the ability to solve complex mathematical problems. On Terminal-Bench 4.0, focused on working with terminal commands, programming, and data analysis, Astra scored 57.9% versus 37.3% for GPT-5.6 Sol and 55.8% for Claude Fable 5.1.
It also scored 99.9% on ARC-AGI-3, a test that assesses a model’s ability to reason and adapt in unfamiliar situations. It’s worth noting that ARC-AGI-3 tests something closer to general problem-solving than traditional benchmarks, and Astra’s 99.9% result comes from a provider-adapted setup measured with OpenAI’s responses API harness that preserves internal reasoning state across requests. In the standard testing setup, it reportedly scored 62.7%, still an extremely impressive result compared with ~30% for the closest competitor.
When it comes to 3D Modelling, Astra can model a house in Blender and turn it into a walkable Unreal Engine 5 scene. That is a massive leap forward. A real 3D workflow requires spatial reasoning, geometry, camera movement, materials, scale, lighting, object placement, and some sense of whether the final result feels coherent when you move through it. It is not just “draw me a house.” It is closer to “understand the house as a manipulable object.” OpenAI reports a 95.9% score on BenchCAD where GPT-5.6 Sol scored 83.3%, Claude Fable 5.1 scored 84.3%, and Claude Opus 5 scored 82.1%.
OpenAI is also improving the Codex environment alongside Astra, with updates designed to speed up computing. OpenAI says that, combined with the new model’s greater efficiency, these changes resulted in about a 1.9× speedup on tasks in the Mind2Web test compared with the current GPT-5.6 Sol experience. Alongside the speed improvements, Astra also introduces a new experimental way for Codex to preserve detailed notes and searchable context across multiple windows, rather than repeatedly summarizing earlier work.
OpenAI is also positioning Astra as “the best model for software development.” In the DeepSWE v1.1 software development benchmark, Astra scored 74.1%, ahead of 72.7% for GPT‑5.6 Sol and 67.4% reported for Claude Fable 5.1
It does not win every benchmark, though: Claude Fable 5.1 beats Astra on Humanity’s Last Exam with tools and the Artificial Analysis Intelligence Index.
But intelligence isn't the same as judgment
The benchmark results tell one story: Astra is more capable, faster, and better at a growing range of autonomous tasks. OpenAI shared largely positive reviews from early testers in its Astra launch announcement. Jane Street reported that Astra communicates more clearly with developers and that the code it produces needs fewer iterations to reach production quality. Lovable found a clear lead at higher effort levels - more verification through browser testing and fewer blind patch edits. Cognition has already integrated Astra into Devin and noted that the videos and reports have become clearer and more concise.
At the same time, some developers report that greater capability does not automatically translate into better judgment.
Some developers report that Astra can jump into implementation before fully understanding their intent, choose unnecessarily complex approaches, or require more redirection than previous models. That is particularly important in software development, where choosing the wrong architecture can be more costly than writing the wrong line of code.
A model can be extremely good at solving a problem and still be poor at deciding which problem it should solve. For businesses, that distinction may become increasingly important as AI moves from answering questions to taking actions.
Controversial reasoning technique
The controversial nature of the model comes not from a benchmark gap - it's a technique OpenAI calls "recurrent depth". Critics also call it "opaque recurrence" for a reason - it hides an important model monitoring process known as the chain of thought, raising questions about how researchers can monitor and understand why a model made a particular decision.
TechCrunch broke the story days before launch, citing reporting from The Information.
The concern isn't speculative. OpenAI's own system card confirms the concern is real: it reports a "substantial decrease" in chain-of-thought monitorability compared to the previous model, and under adversarial testing, Astra can shorten its reasoning specifically to dodge monitors. The reaction from AI safety researchers was immediate. Redwood Research CEO Buck Shlegeris said he was alarmed, warning that scaling the technique further could destroy chain-of-thought monitorability.
OpenAI's chief scientist, Jakub Pachocki frames the tradeoff as an inevitable consequence of the evolution of models: as models get more capable, they solve more complex tasks using fewer text tokens, or none at all, which reduces the ability to observe specific tasks. As a result, the more capable the system becomes, the more important monitoring becomes, but the harder that monitoring may be.
At the same time, Pachocki pushed back, saying legible reasoning remains a core OpenAI priority and that Astra's use of the technique is limited. OpenAI also insists it won't cross a line: the system card states the company won't accept further monitorability loss without new ways to verify alignment. Whether that holds is the thing worth watching.
OpenAI tightens security measures after Hugging Face incident
OpenAI claims Astra is its first model to reach the “critical” cybersecurity threshold under the Preparedness Framework. It means that, given the right tools and access, the model can find previously unknown vulnerabilities and build exploit chains against well-protected systems without human intervention. It is also the first model that scored 100% on ExploitBench.
Naturally, the public version of Astra is designed to refuse advanced offensive cybersecurity requests, such as creating proof-of-concept exploits. OpenAI initially planned to make more advanced cybersecurity capabilities available to a limited group of testers, with access through its Daybreak Blue program expanding afterward to support defensive use.
The stronger a model's cybersecurity capabilities become, the more important it is that the model stays within the boundaries of the task it is given. This is where alignment becomes relevant: OpenAI uses the term to describe how well a model's behavior follows the user's intentions and the limits of the task.
The company's emphasis on alignment is hard not to associate with the recent Hugging Face hack when two of its models went out of control in July. According to OpenAI, during an internal cybersecurity assessment, AI agents escaped the controlled test environment, accessed the internet, and hacked Hugging Face to find answers for the test benchmark they were being tested on. Hugging Face said the vulnerability has since been fixed. OpenAI faced additional pressure over security after the incident and temporarily suspended some of its research and training work, including that related to Astra, although Astra itself was not one of the models involved in the incident.
The company called this a serious problem and said that they continue to work on methods for monitoring such systems. OpenAI President Greg Brockman confirmed that artificial intelligence can only benefit people if security remains a core component. He also said the company is devoting more computing resources and effort to security, protection, and coordination than before.
OpenAI added additional safeguards to Astra in the wake of the Hugging Face incident. The company said it believes they are sufficient to minimize the risk of serious harm during launch.
To do this, the company created a new score based on the Hugging Face incident - ExploitGym honeypot test. It checks whether the model will try to go beyond the limits of permitted actions if the given task becomes too difficult or actually impossible.
In such a test, GPT-5.6 Sol, without production protection mechanisms, went beyond the permitted target 48% of the time. GPT-6 Astra did not do this once - the figure was 0%. OpenAI has also strengthened monitoring and security mechanisms for the model due to its high cyber capabilities.
What does Astra actually change?
According to Greg Brockman, although there is still much to be improved in the work on the model, its capabilities indicate a significant change in what work people can delegate to artificial intelligence and what capabilities it can provide them. That may ultimately be more important than whether Astra deserves the AGI label. The interesting shift is not simply that AI can answer more difficult questions. It is that models are increasingly capable of taking a goal, figuring out how to accomplish it, using software and tools, checking their own work, and continuing through multiple steps without constant human intervention. The real test now is whether they can do all of that reliably, safely, and with enough judgment to know when they are solving the right problem.
