- OpenAI calls GPT-6 Astra its most intelligent and aligned model yet.
- The model can interact with computers and websites to complete multi-step tasks.
- It can research online, fill out forms, create spreadsheets, test websites and troubleshoot software.
What if using AI no longer meant sitting beside it and telling it what to do every few minutes?
That is the bigger question behind GPT-6 Astra, the latest model from OpenAI, which the company calls the world’s most intelligent and aligned AI model.
The appeal is not simply that Astra can produce better answers. It is that OpenAI wants the model to handle tasks, not just respond to prompts. Think about the difference.
You could ask it to research something online, fill out information, work across different websites, create a spreadsheet, test a website or troubleshoot software. Instead of stopping after giving you instructions, Astra is designed to interact with the computer and work through multiple steps to complete the task.
That could change the relationship people have with AI.
From asking questions to handing over jobs
Most AI tools still require a surprisingly large amount of supervision. You ask a question, check the answer, correct the instruction, give another prompt, review the next response and repeat the process.
Astra is designed to move beyond that back-and-forth.
OpenAI says the model can understand an overall objective, work out the steps needed to achieve it and adjust when circumstances change. When something genuinely requires human judgment, it can ask a focused question rather than constantly waiting for instructions.
That matters for the kind of computer work that quietly eats up hours.
An employee might need to gather information from several websites, enter it into a spreadsheet, check the results and prepare a presentation. A developer might need to test a website, identify what went wrong and troubleshoot the problem. Someone handling administrative work could need to move information between systems or fill out online forms.
Astra is designed to operate across those environments rather than treating every step as a separate conversation.
In other words, the goal is less “tell me how to do this” and more “here is what needs to get done.”
The speed question
OpenAI also says Astra performs better on tasks that require sustained computer interaction.
In the company’s OSWorld 2.0 latency simulations, GPT-6 Astra completed tasks in an average of about 40 minutes, compared with roughly 75 minutes for GPT-5.6 Sol. OpenAI says that represents about 47 per cent less time spent completing the simulated tasks.
The company also reports extremely high results on several demanding benchmarks, including 98 per cent on FrontierMath Tier 4, 99.9 per cent on ARC-AGI-3 and 100 per cent on ExploitBench.
Those numbers should be viewed in context: benchmark performance does not automatically translate into flawless performance in everyday work. But they illustrate what OpenAI believes its new model can handle.
And there is another important piece to this.
A system that can operate independently has to know when not to act.
OpenAI says GPT-6 Astra has been trained to better distinguish between instructions, recognize when clarification is genuinely necessary and continue making reasonable assumptions when small gaps do not need human intervention.
That could be just as important as making the model more capable.
Because the dream is not simply to have a smarter chatbot.
It is to have an AI that can take a piece of work, figure out what needs to happen, navigate the tools required and return with a result, without turning the user into a full-time AI supervisor.
That is why GPT-6 Astra could represent a more meaningful change than another jump in chatbot intelligence.
The real test may be whether people eventually stop thinking about AI as something they have to operate and start treating it as something they can delegate to.
OpenAI is initially making GPT-6 Astra available to a limited group of organizations before expanding access to ChatGPT Plus, Pro, Business and Enterprise users. Developers will also be able to access the model through the OpenAI API.
