OpenAI released GPT-6 Astra on September 3. The line that travelled fastest came from OpenAI President Greg Brockman, who ended the press briefing with: “Welcome to the AGI era.”

Brockman said he personally believes we may now be in that era. I would be more careful.

AGI has never had one clean test. Even OpenAI and outside researchers use different definitions. A model can beat humans at a benchmark and still fail at things a normal person handles without thinking twice.

What interests me more is something much more practical.

Astra is becoming very good at using the computer itself.

One Topic: One Topic: GPT-6 Astra and 7 Ways to Use It

GPT-6 Astra – The bigger change is computer use

Previous models could tell us what to do, do it as well up to certain extend. Astra is much better at doing the work inside browsers and computer.

OpenAI shows it filling forms, updating CRM records, organising calendars, researching online, working inside document editors, analysing data, building websites, testing interfaces and troubleshooting software.

On OSWorld 2.0, a benchmark for operating computer interfaces, Astra scored 72.6% versus 65.7% for GPT-5.6 Sol. OpenAI says its simulated completion time dropped from roughly 75 minutes to 40 minutes.

That difference matters more to me than another chatbot benchmark.

AI is moving from “give me an answer” toward “give me an outcome.”

But has GPT-6 Astra actually reached AGI?

The strongest evidence is also the part that needs the most context.

OpenAI reports 99.9% on ARC AGI 3. ARC Prize tested the same model with its standard setup and got 62.7%. Both are record results, but the difference comes partly from how Astra can preserve reasoning state and context between attempts.

Even more interesting, Astra used fewer actions than the median human on 96% of the tested levels.

ARC Prize still says clearly that saturating ARC AGI 3 would not prove AGI. Artificial Analysis has also put Claude Fable 5.1 ahead of Astra on its newly updated broader Intelligence Index.

So I am not ready to call this AGI.

7 ways you can actually use GPT-6 Astra

1. Give GPT-6 Astra a browser job

Instead of asking, “How should I clean up my CRM?”, give it the actual outcome.

Ask it to identify duplicate records, correct obvious formatting problems, update fields according to your rules and prepare a list of anything uncertain for your review.

The browser becomes part of the task rather than something you operate between AI conversations.

Start with work you can easily reverse.

2. Turn a pile of files into a decision

Astra’s API supports a 1,050,000 token context window.

That makes it useful for work where the problem is scattered across reports, spreadsheets, PDFs, notes and earlier decisions.

I would give it the source material plus one clear decision: “Should we renew this supplier?”, “Which project needs intervention?” or “What changed from the original business case?”

Then require every conclusion to point back to evidence.

3. Audit a spreadsheet before you send it

This is one of the uses I would try quickly.

Give GPT-6 Astra the spreadsheet, the business rules and the previous approved version. Ask it to check formulas, totals, mismatched assumptions, missing values and unusual movements.

Then ask for a clean management summary of what changed.

OpenAI has trained Astra specifically for professional documents, spreadsheets and presentations, including following existing templates.

4. Build a website and test it

ChatGPT Sites already lets users create and host websites and web apps from a prompt.

GPT-6 Astra adds stronger visual judgment and computer use. So the workflow can continue after the first page appears.

Give it your requirements, let it build the site, then ask it to open the result, click through every important path, test forms, check different screen sizes and repair what fails.

That last part is what makes this more useful than another website generator.

5. Take a coding task all the way to verification

For developers, I would stop asking only for code snippets.

Give Codex the issue, repository, acceptance criteria and tests. Let Astra inspect the codebase, make the change, run tests, open the application, check the result and prepare the pull request.

GPT-6 Astra can also preserve notes across long Codex sessions and search earlier context when a project runs beyond one context window.

6. Let it operate specialist software

This might become Astra’s most interesting use.

OpenAI has shown it working across tools such as Power BI, engineering software and visual applications. Early tests shared on X also show Astra taking loosely defined work inside tools such as Blender and beginning the task without needing every click explained.

Many businesses still depend on software with poor APIs.

An AI that can use the interface itself changes what can be automated.

7. Hand over the annoying life admin

Not every use needs to be enterprise work.

OpenAI demonstrated Astra handling tasks around apartment searches, appointments, product research and other web based errands.

A useful instruction might be: find three suitable options under my constraints, compare them, complete everything you can safely complete, then stop before any payment or commitment.

That is where I would start personally.

A contained task. Clear boundaries. Easy to check.

My rule for using GPT-6 Astra

I would not start with the most sensitive workflow I have.

Give it a job where success is obvious and mistakes are reversible. Define what it can change, what needs approval and what it must never touch. Review the first few runs.

If it earns more responsibility, give it more.

That is a much more useful test than arguing over whether we have officially crossed the AGI line.

---

Interested in travel or photography, read last week’s LensLetter newsletter about Luminar big 2026 update.

Read last week’s JustDraft about task you can give to personal AI agent.

---

Two Quotes to Inspire

Ambition gives you more options. Strategy gives you fewer things to chase.

Leaders who chase the headline number always end up managing the footnote.

---

One prompt to steal

Use this when you want GPT-6 Astra or another computer using agent to finish work instead of giving advice.

Complete this outcome:

[Describe the finished result you want]

Use these sources and tools:
[List the files, websites, apps or folders it may use]

You may:
[List the actions it can take without asking]

You must stop and ask me before:
Sending messages
Publishing anything
Deleting information
Making a purchase
Changing permissions
Making an irreversible decision

Do not interrupt me for routine choices.
Make reasonable assumptions when the consequence is small.

Keep a short record of the important actions you take.

Before finishing, verify the result against these success criteria:
[Add your criteria]

If verification fails, identify the cause and try one different approach.

At the end, give me:
1. The finished result
2. What you changed
3. Assumptions you made
4. Anything still requiring my approval

The permission section is the part I would not skip. A smarter agent deserves clearer boundaries, not looser ones.