GPT-6 Astra shipped late last week. It’s a different model. So I read what OpenAI published, pulled what its product lead said on X, and ran 100s of my own tests. Here’s what’s worth doing in your first week.
Most AI marketing stack advice argues about which tool to buy. This is the other half: what to change once it’s in your hands.
GPT-6 Astra on low performs better than GPT-5.6 Sol on high.
The 10 tips, at a glance
- Check which Astra your plan actually gives you, because Plus doesn’t get it in regular chat.
- Turn the reasoning effort down, not up. It has a different curve.
- Learn your usage meter before you need it. There are two limits, and you need room in both.
- Feed it context once, properly, instead of re-explaining yourself every session.
- Turn on computer use on purpose, this is the most impressive thing I’ve ever seen in a browser.
- Write assignments using OpenAI’s own four-part frame.
- Plan first, then set the goal, for anything that takes more than one step.
- Paste OpenAI’s anti-slop word list into your custom instructions. This is the best five minutes you’ll spend.
- Stop the permission loop with the prompt OpenAI published for exactly that problem.
- Watch the 272K context cliff.
1. Check which Astra your plan actually gives you
The name changes by plan. That’s where most “I don’t have Astra” confusion comes from.
In ChatGPT chat, Astra shows up as GPT-6 Pro, and only on Pro $100, Pro $200, Business and Enterprise. Enterprise needs an admin to switch it on. Plus includes Astra in ChatGPT Work and Codex, but not GPT-6 Pro in Chat.
Codex CLI needs 0.153 or newer, and the desktop app may need updating twice. Menu, Check for Updates, quit fully, reopen, check again. Even if it just auto-updated.
2. Turn the reasoning effort down, not up
This tip runs against all your prior AI instinct.
Tibo Sottiaux, who leads core products at OpenAI, said on September 6 to more than 22,000 likes:
To calibrate you all on which reasoning effort to use for Astra, know that GPT-6 Astra on low performs better than GPT-5.6 Sol on high. If you were using high reasoning efforts with Sol and were happy, I suggest you move down to low or medium for Astra.
I’ve confirmed this myself through extensive testing. Astra on low is a beast.
3. Learn your usage meter before you need it
Astra is metered differently than anything you’ve used, and there are two clocks running.
Limits apply over a five-hour window and a weekly window, and you need allowance left in both to keep working. A new five-hour window starts when you send your first message after the last one ends. Settings, then Usage, shows where you stand and when it resets. OpenAI’s own estimate of local messages per five-hour period (as of September 9):
| Model | Plus Chats | Pro 20x |
|---|---|---|
| GPT-6 Astra | 5 to 45 | 100 to 900 |
| GPT-5.6 Luna | 250 to 2,000 | 5,000 to 40,000 |
Five messages is a real limitation. Do your thinking with a cheaper model and bring Astra in for the hard part.
There is talk they may limit new Pro accounts so learn to trade down or get Claude, Gemini or Grok in your workflows as your second model.
4. Feed it your business once
Astra holds a 1,050,000-token context window and takes up to 922,000 tokens of input. On Business and Enterprise, an admin installs Company Knowledge, connects your sources, and then you pull it in with @Company Knowledge or the tools menu. On Plus and Pro, the equivalent lives in Settings under Personalization, plus Projects for anything ongoing.
This is the #1 thing I see folks miss. The right context means more than the right model in the vast majority of knowledge work.
5. Turn on computer use on purpose
Astra can drive your actual computer.
It’s desktop-app only. Plugins, then Computer Use, then Install. On a Mac you’ll grant Screen Recording and Accessibility. You call it with @Computer or by naming an app. On Windows it takes over your foreground, so you’re not doing anything else while it runs.
It acts inside browser sessions you’re already logged into. That’s why it’s useful for pulling numbers out of ad platforms, and why I watch it like a hawk.
6. Write assignments
OpenAI publishes a four-part framework:
- Goal: what should it do?
- Context: what information or sources help?
- Output: what format, length and level of detail?
- Boundaries: what must stay unchanged, and what should it check with you before doing?
You don’t need all four every time. Boundaries is the one marketers skip and then regret, because it’s where QA happens. If you don’t want AI slop, the delta is in how clearly you set your boundaries / QA.
7. Plan first, then set the goal
For anything multi-step, run /plan, then convert it with /goal. The goal text becomes both the instruction and the definition of done, which stops a long task from wandering.
Enter steers it mid-task. Tab queues your message for the next step. Knowing which one you just hit saves a lot of confusion.
The best marketing artifact of this launch is a word list OpenAI published for developers.
8. Paste OpenAI’s anti-slop list into your custom instructions
OpenAI shipped a literal blocklist of the words that make AI writing sound like AI writing. It’s buried in developer docs, so nobody in marketing has picked it up. Paste this into your custom instructions:
Avoid using slop words or phrases like “Bottom Line:” in conclusions, “delve,” “foster,” “leverage,” “it’s worth noting,” “importantly,” “Question? Answer.” or “This isn’t about X. It’s about Y.”, “genuinely” or hyphenated compound descriptions and adjectives. Do not use concluding summary statements such as “In short:..”, “The simplest mental model is:…”.
State the intended action directly. Avoid adding what you won’t do, what will remain unchanged, or how you’ll separate or categorize results. Do not use contrastive framing such as “X, not Y” that introduces an unprompted alternative that the user didn’t ask about. Avoid invented compound labels, vague qualifiers, and canned transitions; use plain verbs and prepositions to state the actual relationship directly.
Read that second paragraph again. Those are the exact tics that make a LinkedIn post smell generated. Works on any model.
9. Stop the permission loop
Astra asks more questions than the last model. OpenAI says: it’s “more likely to ask the user a question when additional input could materially change the result,” which means it stops where you expected it to just get on with it.
They published the fix. Add this:
Before asking the user clarifying questions, you should complete the work that is already authorized from context and necessary to make the proposed action concrete and reviewable. The user should be approving a concrete, reviewable result.
Same page, and it matters if you’ve built up custom instructions over time: Astra is more sensitive to instruction files, and OpenAI strongly recommends auditing the ones your model can read. Old instructions that were harmless before can stall work now.
10. Watch the 272K cliff
Through the API, standard pricing is $10 per million input tokens and $50 per million output, with cached input at $1.
Then there’s a cliff. Prompts above 272,000 input tokens re-price the entire request at 2x input and 1.5x output. Not the overage, the whole thing. A million-token context is available, but it will cost you.
What not to believe yet
The model is a week old, so the advice market is running well ahead of what I can prove via testing. Three claims I’d skip:
- “Higher effort actually uses less quota.” A popular thread argues extra-high one-shots cheaply while medium burns tokens on retries. OpenAI’s documentation says the opposite, and my tests disagree. Unresolved.
- “Delete your instruction files.” A video with 105,000 views recommends this. Don’t do this. Anytime there is a new model have that model edit them to its taste.
- Benchmarks from before September 6. OpenAI shipped an efficiency change that Sottiaux said can draw “up to 3-4X less usage,” and reset limits repeatedly during launch week. Numbers measured before that are for a product that no longer exists.
The one I’d do today 👇🏼
If you only do one, do number 8. It costs five minutes, it works on every model you own, and it fixes the thing your boss or clients will notice.
Sources
All checked on September 9, 2026.
- GPT-6 Astra model page: pricing, context window, the 272K threshold
- Using the latest model: the anti-slop prompt, the initiative prompt, the instruction-file warning
- Prompting: Goal, Context, Output, Boundaries
- GPT-5.6 and GPT-6 Pro in ChatGPT: plan mapping, Codex CLI version, the double update
- Managing usage with GPT-6 Astra: the two windows and the message estimates
- Tibo Sottiaux on reasoning effort and on the usage efficiency change