GPT-6 Astra Targets the Office Work That Pays Your Salary

In Brief

OpenAI's GPT-6 Astra puts artificial intelligence to work inside the software people use every day. 

Its launch sharpens a difficult question for business leaders: How do you prepare employees when the technology is advancing faster than the organization can absorb it?

What Happened

OpenAI introduced GPT-6 Astra on Sept. 3, with computer use at the center of its pitch. That means AI can operate applications, click through screens, and complete tasks across software, and operate a computer for days at a time. OpenAI's examples include Excel work, Power BI dashboards, and presentations that follow company templates.

The company reports higher scores and faster performance on OSWorld 2.0, an evaluation of computer tasks. Those results come from controlled testing, including time simulations. They don't establish that Astra can reliably handle every company's work. OpenAI announced a phased rollout across paid ChatGPT plans and its developer platform.

SmarterX founder and CEO did not have access when Episode 237 of The Artificial Intelligence Show was recorded on Sept. 4. But Roetzer's reaction was based on the launch materials, safety disclosures, and early reports. What concerned him most was the distance between the work these systems are beginning to perform and the preparation happening inside organizations.

The Key Numbers

72.6% - Astra's OpenAI-reported partial score on OSWorld 2.0's offline tasks

65.7% - GPT-5.6 Sol's score on the same evaluation

47% - Reduction in time per task in OpenAI's OSWorld simulations

97.6% - Astra's score on FrontierMath's hardest tier of research-level math problems, as reported by OpenAI

Computer Use Exposes the Readiness Gap

The examples are recognizable paid work. Spreadsheets, dashboards, and presentations occupy real hours in real jobs. For an employee watching those demonstrations, the questions extend to which parts of their role will change and who will help them adapt. Roetzer describes an unusually uneasy response: "This is the first model release where my first actual reaction was dread over excitement," he says.

Many organizations are still learning the previous tools. "Traditional organizations are still trying to absorb AI assistant capabilities from 2024," Roetzer saya. His concern is that companies built around AI will move quickly with agents, systems that can complete multiple steps to finish work, while established businesses are still teaching employees basic prompting or single-step assistant use. That creates competitive pressure before a workforce has had time to understand the change.

More capability also raises harder oversight questions. Roetzer also focused on what happens when people have less visibility into how an advanced artificial intelligence model reaches a decision. That concern grows when the system can act inside workplace software. OpenAI says it delayed parts of Astra's development and release while strengthening protections against cyber misuse and unauthorized actions. For Roetzer, the urgency is clear: "The technology is racing ahead in its capabilities."

"I find myself stuck in the middle of understanding the technology, advocating for it responsibly and in a human-centered way, but totally empathizing with people who don't want it and don't like it."

— Paul Roetzer, founder and CEO of SmarterX

SmarterX Take

A computer-use demonstration should start a conversation with the people who do that work. Ask them which steps are repetitive, where errors become costly, and what a finished result needs to look like. Their knowledge is essential to judging whether delegation actually helps.

Leaders also need to explain what they intend to do with any time saved. Employees deserve a clearer answer than a launch video's promise of higher productivity. A successful trial should produce evidence about quality and oversight, alongside an honest discussion of how responsibilities change. A benchmark score cannot answer those organizational questions.

What to Watch

Every-day reliability will determine the value. As access expands, the useful evidence will be completed work with acceptable quality, reasonable costs, and manageable human review. A faster attempt has limited value if an employee must reconstruct it afterward.

Will workers accept and adopt advanced models. Roetzer expects growing resistance when people see what computer use means for their work. Access to a stronger model does little to resolve employees’ concerns about how their roles will change.

How Ready Are Organizations for Computer-Using AI?

Only 25% of respondents say their organization has reached the Scaling phase of AI adoption, according to the 2026 State of AI for Business Report. That phase means broad adoption with consistent improvements in efficiency and performance. The finding puts Astra's launch in context: a more capable tool arrives in a business environment still building the ability to use AI consistently.

Based on more than 2,100 responses across roles, functions, and industries, the report examines adoption, training, governance, and workforce sentiment. It gives leaders a basis for assessing the preparation their own organizations need as AI takes on more work. Read the full report →