The large language model (LLMs) race has never been this fierce.
What started after OpenAI introduced ChatGPT, the moment turned AI development into a cycle of increasingly frequent releases, bigger training runs, and increasingly ambitious claims about what software can actually do.
OpenAI, Anthropic, Google, Meta and others are competing at a pace that makes each new generation feel less like a distant technological milestone and more like the next move in an ongoing contest.
That acceleration is becoming especially visible as models move beyond generating answers and toward carrying out work.
And this time, OpenAI is trying to leap ahead with the release of 'GPT-6 Astra.'
This model, which OpenAI announced as the latest frontier model, is presented as a major shift toward systems that can operate across computers, browsers, codebases and professional software rather than simply responding to prompts.
The company describes it as its most capable and aligned model so far, with state-of-the-art performance in computer use, browsing, software engineering, cybersecurity, science and professional work.
The most important change is therefore not simply that Astra produces better answers.
OpenAI is positioning it as a system that can take a goal, work through multiple steps and interact with the tools needed to complete it. The company gives examples ranging from filling out online forms and updating CRM records to researching information, working in document editors, creating websites, testing software and troubleshooting problems visible on a user's screen.
In OpenAI's own latency testing on OSWorld 2.0, Astra achieved a higher score than the powerful and capable GPT-5.6 Sol while taking roughly 40 minutes per task compared with about 75 minutes for Sol.
That emphasis on computer use also explains why the model is being framed as something closer to an agent than a conventional chatbot.
Astra can continue working through multistep processes, adapt when instructions change and handle tasks that span different applications.
OpenAI says it has also redesigned aspects of Codex around the model, including a system that can preserve and retrieve context from earlier windows rather than repeatedly compressing long sessions into summaries that may lose important details.
The headline benchmark figures are striking. OpenAI reports a 98% score on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench, while also claiming state-of-the-art results on Agents' Last Exam, AutomationBench and ScreenSpot Pro.
The ARC Prize Foundation separately reported that Astra reached 99.9% on ARC-AGI-3 under a provider-adapter setup and said the model used fewer actions than the median human on 96% of tested levels.
At the same time, independent discussions have pointed out that some of these results depend heavily on the evaluation harness and that direct comparisons across benchmarks are not always straightforward.
The coding and software side of the release may ultimately matter more for everyday use.
OpenAI says Astra is its strongest software engineering model to date, with improvements in code generation, browser testing, debugging and longer-running agentic work.
Its API documentation lists a 1.05 million token context window and up to 128,000 output tokens, while allowing reasoning effort to be adjusted across several levels. That combination is aimed less at isolated questions and more at keeping a large project in context while the model works through it.
The cybersecurity results are even more consequential, because they show where capability and risk are beginning to overlap.
OpenAI says Astra is its first model to reach the Critical level for cybersecurity under its Preparedness Framework.
In testing, it achieved 100% on ExploitBench and 88% on SRE-Bench in a single attempt, rising to 99.2% after four attempts.
OpenAI also says that during evaluations Astra discovered and used two previously unknown zero-day vulnerabilities, which the company is disclosing to the relevant maintainers.
That is why the safety component of this release is receiving almost as much attention as the model itself.
OpenAI says it strengthened monitoring, isolation, access controls and model safeguards because Astra's capabilities create new ways for both users and the model itself to cause harm.
The company's system card also describes an important limitation: in some adversarial evaluations, Astra showed evasive behavior when it was aware of monitoring, although full-context monitoring was able to detect the tested attacks.
The model is therefore being deployed with restrictions around its most advanced offensive cybersecurity abilities rather than exposing all of them immediately.
The timing of Astra's arrival adds another layer to the story.
OpenAI's own recent research updates have centered on the security consequences of increasingly capable agents, including its August investigation into an incident involving Hugging Face.
The Astra launch comes amid a broader period of rapid movement from competing AI companies, making the race increasingly about who can build systems that not only reason well, but can reliably act in the real world.
OpenAI has also attached unusually strong language to the release.
In a closing remark at the press briefing, President Greg Brockman described Astra as a potential beginning of the "AGI era", while OpenAI's launch material says the model represents a new generation of intelligence.
That claim is likely to remain contested because AGI has no universally accepted benchmark or definition, and much of the evidence for Astra's significance still comes from company-selected evaluations.
What is less ambiguous is the direction of development.
The frontier models are increasingly being built to execute long sequences of actions, work across software and recover from problems instead of simply answering a user and stopping.
Astra is initially rolling out to a limited group of organizations, with OpenAI saying access will expand to Plus, Pro, Business and Enterprise users as well as through its API and AWS.
The rollout itself has already become part of the conversation, with users on social platforms reacting to the gap between the announcement and broader availability, while others are debating whether the benchmark results represent a genuine step toward AGI or simply another iteration in an unusually fast-moving model race.
The significance of GPT-6 Astra may therefore depend less on whether it can be called AGI and more on what happens when people begin using it for sustained, unsupervised work.
The interesting question is no longer only whether a model can solve a problem.
It is whether the model can understand a larger objective, decide what needs to happen next, use the necessary tools, recognize when something has gone wrong and finish the job without constant intervention.
Astra is OpenAI's clearest attempt yet to move that idea from demonstration toward a general computing capability.
