OpenAI released GPT-6 Astra on September 3. Access began with a limited group of organizations, then opened to paid ChatGPT plans and API customers within days.
For marketers, Astra's biggest change is what it can do outside the chat window. The model can operate a computer directly, looking at the screen and controlling the mouse and keyboard much like a person would.
OpenAI calls this capability computer use. In practice, Astra takes a screenshot of whatever is on screen, decides where to click and what to type based on what it sees, performs that action, then takes another screenshot to check what actually happened before deciding on the next step. It repeats that loop, click by click, until the task is finished. That means it can work inside Google Ads, Salesforce, a spreadsheet, or other software through the interface itself, without needing a custom integration built for it first.
How Astra Builds on Earlier Computer-Use Agents
Anthropic released Claude's computer use in October 2024 and admitted it was slow and error-prone. OpenAI followed with Operator in January 2025, then folded it into ChatGPT agent mode that July. Perplexity brought similar capability to Comet.
Those early versions lost their place, clicked the wrong button, or stalled halfway through a task. Astra arrives after nearly two years of the whole industry trying to make computer use reliable enough to actually trust with real work.
GPT-6 Astra Is Faster at Computer-Use Tasks
OpenAI says Astra completed computer-use tasks in roughly 47 percent less time than GPT-5.6 Sol, while scoring higher. In OSWorld 2.0 simulations, Astra averaged about 40 minutes per task versus 75 minutes for Sol. Those numbers are OpenAI's own.
Our internal testing backed that up. Astra stayed on task through longer sequences, recovered from small mistakes on its own, and needed far fewer corrections along the way. Work that used to feel like babysitting a confused intern started to feel more like handing something off to a competent assistant.
It also takes new instructions mid-task without losing the original assignment. You can correct a date or change a requirement while it's already working, and it adjusts instead of treating the interruption as a brand-new request. Earlier agents could not do that reliably.
What GPT-6 Astra Can Do for Marketing Teams
Marketing teams live across ad platforms, analytics tools, CRMs, spreadsheets, presentation software and email, and most of those tools were never built to talk to each other. Astra can move through that stack itself, even when the software doesn't have an AI integration built in.
For marketers, that could mean checking campaign settings against a brief, pulling numbers into a report, monitoring competitor pages, organizing research or preparing a first draft. You still set the goal and judge the result. Astra handles more of the screen work in between.
The Risks of Giving Astra Access to Marketing Tools
Giving Astra that level of access comes with obvious risks. A wrong click can change a budget, publish an ad, send something private or delete data that mattered. OpenAI itself recommends limiting what Astra can touch, requiring approval on anything consequential, and actually checking the result rather than trusting the summary.
Astra also crosses the Critical cybersecurity threshold under OpenAI's Preparedness Framework, the first model from the company to do so. During testing, Astra built working exploits from known vulnerabilities, discovered previously unknown vulnerabilities and turned some into working exploit chains. The rollout is staged, with advanced cyber requests restricted in the public version. For marketers, the caution here is pretty clear: the same model you might point at your ad accounts is powerful enough that OpenAI built additional safeguards around how it can be used. Worth keeping your own guardrails tight.
The practical starting point is small. Pick one repeatable task, narrow what Astra has permission to touch, and put an approval step anywhere money, customer data or publishing is involved. Track how accurate it actually is before you hand over anything bigger.
Better Execution Still Depends on Better Decisions
Astra pushes an existing shift in marketing much further. As agents become capable of handling more of the work between a request and a finished task, execution itself becomes less differentiating.
Marketers still choose which bets are worth making: which audiences matter, which campaigns to run and which to leave alone.
An agent working from screenshots can execute the decision faster, but it doesn't necessarily make the decision better.
That's part of the problem Lickly was built to address. RAD Intel's Decision Intelligence company starts with the audience a brand needs to reach, then helps marketers evaluate where budget should go before it's committed and understand the reasoning behind the recommendation.
Computer use is moving quickly from demo to practical use, and Astra suggests that shift will continue. Moving first won't help much if you're automating the wrong decisions. Marketers still need to know what to hand over, what requires human judgment and why the decision made sense in the first place.
