8 Things You Should Try with GPT-6 Astra in ChatGPT

OpenAI released GPT-6 Astra on 3 September 2026 and calls it its most capable model so far. The headline isn't that it writes better paragraphs. It's that Astra can do things: work inside apps, browse, build spreadsheets and finish long tasks without you steering every step.

That changes how you should use it. If you only ask Astra the questions you asked older models, you'll barely notice the upgrade. Here are eight things worth trying first.

Who has it

Plan	Astra?
Free	No (GPT-5.6 Luna)
Plus ($20/month)	Yes, rolled out gradually from September
Pro	Yes, with more usage and a higher-effort Pro mode
Business / Enterprise	Yes. Enterprise admins must switch it on

Usage counts against your normal plan limits. In the API it's gpt-6-astra at $10 per million input tokens and $50 per million output tokens.

1. Hand it a form-filling chore

Astra's biggest jump is computer use: it can click, type and move between applications, even ones without an API. On OSWorld 2.0, a benchmark of real desktop tasks, OpenAI reports 72.6% for Astra versus 65.7% for its predecessor.

Try a task you'd normally do by hand: "Copy these 15 leads from the attached spreadsheet into our CRM, one record each, and tell me which ones already existed." Watch the first run. Stay in the loop for anything that sends, buys or deletes.

2. Ask for research, not answers

Old habit: "What's the best accounting software for a small agency?" Better with Astra: "Compare five accounting tools for a 10-person agency in India. Check each vendor's current pricing page, note GST invoice support, and give me a table with links."

Astra browses, checks sources and builds the comparison. Ask for links every time, then open the two or three that will decide your choice.

3. Build a real spreadsheet model

OpenAI says Astra can complete Financial Modeling World Cup challenges about four times faster than the winning human. You don't need to be a finance professional to use that.

Try: "Build a 12-month cash-flow model for a small online store with these assumptions. Put assumptions on their own sheet, use formulas rather than typed numbers, and add a sheet showing what happens if sales drop 20%." Then change an assumption yourself and check that the totals update. That's the test of a real model rather than a pretty table.

4. Teach it your house style once

Astra was trained to follow company voice, templates and design standards closely. Give it two or three of your best past documents and say: "Write the next proposal in exactly this structure and tone. Keep our section headings."

It's like giving a new team member your best examples on day one instead of correcting their drafts for a month.

5. Give it a long, multi-step task

Astra holds context better through long, complex work, and OpenAI says it was trained to finish tasks with fewer tokens and fewer retries. So try something you'd normally split into ten prompts:

"Read these three quarterly reports, pull out revenue, margin and headcount, chart the trend, write a one-page summary for the board, and list the three questions the board is most likely to ask."

6. Code with it in Codex

On Terminal-Bench 4.0, a test of command-line coding work, OpenAI reports 57.9% for Astra, ahead of its previous model at 37.3%. In Codex, an experimental feature lets Astra take notes across long sessions and search earlier messages and tool outputs. You switch it on in Codex's config.toml for now; OpenAI plans to make it the default.

A good first task: "Find why this test fails intermittently, fix it, and add a test that would have caught it."

7. Use it as a second pair of eyes on code

Code-review company CodeRabbit reported Astra caught about 20% more bugs than its baseline. Paste a pull request diff and ask: "Review this as a senior engineer. List real bugs first, with line references, then risky patterns. Skip style comments."

The "skip style comments" line matters. It keeps the review focused on things that break.

8. Ask it to check its own claims

Box reported Astra was more than 10% less likely to make confidently wrong statements. Lower isn't zero. End important prompts with: "Mark anything you're not sure of, and give me a source for every number."

Then actually spot-check the numbers that matter. It takes two minutes and saves embarrassing corrections.

A word on safety

Astra is the first OpenAI model rated at the "Critical" level for cybersecurity capability under OpenAI's own framework, so OpenAI added extra safeguards. The released model refuses advanced cyber tasks, and business admins can require approval before consequential actions and restrict which websites and apps it can touch. For personal use, the same principle applies: let it draft and prepare, and keep the final "send" or "pay" click for yourself.

In one line: Astra's upgrade is doing rather than talking, so try it on chores, research, spreadsheets and long tasks, and keep a human on the actions that matter.