OpenAI has launched GPT-6 Astra, a new frontier model designed to do far more than produce answers inside a chat window. Announced on September 3, 2026, Astra can operate computers, browse websites, work across software, write and test code, and carry out multi-step professional tasks. OpenAI says the model can fill forms, update CRM records, organise calendars, prepare documents, build spreadsheets and even run website checks.
The launch pushes the AI race toward agents that can finish digital work rather than simply explain how to do it. Astra is initially reaching a limited group of organisations, with access for ChatGPT Plus, Pro, Business and Enterprise users, the API and AWS planned over the following days.
What Makes GPT-6 Astra Different From Earlier ChatGPT Models?
The biggest change is action. According to OpenAI’s GPT-6 Astra announcement, the model can interact with computer interfaces and move through workflows involving several tools. A user could ask it to research apartments, organise results, prepare a comparison and continue through relevant websites instead of stopping after producing suggestions.
OpenAI also says Astra can install and test software, troubleshoot what appears on screen, create websites, analyse scientific data and produce polished presentations. Its GPT-6 Astra developer documentation lists a 1.05-million-token context window and maximum output of 128,000 tokens, giving it room to handle unusually large projects and long-running instructions.
That puts Astra closer to a digital worker than a conventional chatbot, although access permissions and human approval remain important whenever the model is allowed to act outside the conversation.
How Strong Is GPT-6 Astra At Computer Use?
OpenAI’s published tests show a sizeable jump on computer-use and automation benchmarks. On OSWorld 2.0, Astra scored 72.6%, compared with 65.7% for GPT-5.6 Sol. OpenAI says the Astra run also completed comparable tasks in about 40 minutes, versus roughly 75 minutes for GPT-5.6 Sol in its latency simulation.
On AutomationBench, which tests professional workflow automation, Astra scored 41.4%, more than double GPT-5.6 Sol’s 18.1%. Its ScreenSpot-Pro result reached 92.7%, according to the same OpenAI benchmark tables.
Key points from the launch include:
- Astra can browse, use software and complete multi-step computer workflows.
- It can create documents, spreadsheets, presentations, websites and code.
- OpenAI is rolling access out first to selected organisations.
- Plus, Pro, Business and Enterprise access is expected to follow.
- API support includes web search, file search, computer use and other tools.
These results do not mean every task will work perfectly. Benchmarks are controlled tests, while live websites can change, block automation or require a person to approve sensitive actions.
Why Is OpenAI Putting So Much Attention On Astra’s Safety?
Computer-use models bring a different risk profile because mistakes can become actions. OpenAI’s GPT-6 Astra safety overview says the model is the company’s first to reach the “Critical” cybersecurity capability level under its Preparedness Framework. With the right tools and access, OpenAI says Astra could discover previously unknown vulnerabilities and develop exploitation methods against well-protected systems.
That capability explains the tighter rollout. OpenAI says it strengthened isolation, monitoring and security controls around the model. Its ChatGPT release notes also say some agent sessions may be paused when monitoring detects a possible misunderstanding of the user’s instructions.
The safety debate is already shaping coverage. Reuters reported on Astra’s launch amid growing scrutiny over autonomous agents, while OpenAI faces the difficult job of increasing capability without losing reliable oversight.
Can GPT-6 Astra Really Complete Entire Tasks For Users?
For many digital jobs, Astra appears built to attempt exactly that. Instead of asking for travel research, then manually opening ten tabs, copying prices and preparing a document, users could increasingly delegate the workflow itself. Similar possibilities exist for sales operations, coding, research, administrative work and data analysis.
Still, “complete the task” should not be read as “operate without limits.” Financial transactions, account changes, cybersecurity work and other sensitive actions may need permissions, confirmations or restrictions. Reliability also depends on the websites, software and data Astra can access.
OpenAI President Greg Brockman has framed Astra as a possible marker of the “AGI era,” a claim that has quickly become part of the launch discussion. Whether that label sticks is less important than the product shift now visible: frontier AI is moving from answering prompts toward navigating software and producing finished work.
FAQs
What Is GPT-6 Astra?
GPT-6 Astra is OpenAI’s newest model for reasoning, computer use, coding, research and professional workflows.
When Did OpenAI Launch GPT-6 Astra?
OpenAI announced GPT-6 Astra on September 3, 2026, beginning with access for selected organisations first.
Can GPT-6 Astra Control A Computer?
Yes, Astra can navigate interfaces, fill forms, use software and perform approved computer tasks autonomously.
Who Can Access GPT-6 Astra?
Selected organisations receive it first, followed by Plus, Pro, Business, Enterprise, API and AWS users.
Is GPT-6 Astra Safer Than Earlier Models?
OpenAI says Astra adds stronger alignment, monitoring and security controls for advanced cybersecurity capabilities overall.
