Welcome back to the AI Edge (by Miles Deutscher) weekly newsletter!
OpenAI released GPT-6 Astra last week, its new flagship model for coding, research, computer use, science, and professional work.
Astra is extremely useful for bigger tasks. You can give it a goal, let it research or work through software, steer it while it runs, and have it return a finished file or result rather than another answer you still need to act on.
In this publication, we’re giving you our ultimate Astra guide so you can get the most out of it.
We’ll cover:
How to Effectively Prompt GPT-6 Astra
How to Cut Token Spend
Computer Use 101
3 Real Astra Workflows (prompts included)
Access is still rolling out, so you may see Astra inside Work or Codex before it appears elsewhere in ChatGPT.
After our main anchor piece, be sure to stick around for the Midweek Edge for all the latest AI news updates, and our Looking Ahead section, where we compile our top AI research.
Let’s get right into things!
Table of Contents
How to Effectively Prompt GPT-6 Astra
Astra works best when you define the result clearly and give it enough room to handle the steps in between.
That means your prompt should give it enough context to make good decisions without scripting every action for the model yourself.
Here are a few tips we recommend (sourced from OpenAI themselves):
Effort Level Selection
Astra has five reasoning levels:
→ Low
→ Medium
→ High
→ Xhigh
→ Max
We recommend keeping Medium as your default.
Move to High or Xhigh for harder research, coding, analysis, or planning. You should reserve Max for problems where the extra reasoning actually matters because it uses more tokens.
Give it the full scope
For longer work, tell Astra what you’re trying to accomplish, what context matters, what it can decide itself, and exactly what finished looks like.
Use this prompting template:
Goal:
[What you want completed]
Context:
[What Astra needs to understand]
Boundaries:
[What Astra can decide itself and what needs your approval]
Done when:
[The exact result you expect before the task is finished]
For example:
Goal:
Research the AI coding market and produce a finished competitor brief.
Context:
I’m using this to decide which products we should study for a new developer tool.
Boundaries:
Make routine research decisions yourself. Ask me before changing the scope.
Done when:
You have compared the major products, pricing, positioning, strengths and weaknesses, then produced a one-page recommendation with sources.
Give it room to “run”
OpenAI recommends letting the model “run". Give the model enough context, and it let it figure out the best execution path.
Make reasonable assumptions for routine decisions and continue working.
Ask me only when missing information would materially change the result, create additional cost, or require an external action I should approve.Clean up old Skill files & instructions
Astra pays close attention to project instructions, Skills, and context files such as AGENTS.md.
Old rules that made sense for another model can interfere with the current task, so review the files attached to important projects and remove instructions you no longer need.
This is also a good time to simplify any giant Skills you’ve built over the last few months.
Steer it while it works
One of Astra’s most useful features is mid-task steering.
If the task is already running and you want to change something, you can update the instruction without restarting the entire job.
Prompting example:
Keep everything you’ve completed so far.
For the remaining research, focus more heavily on enterprise pricing and security, then update the final comparison around those two factors.How to Cut Token Spend With Astra
Astra costs $10 per million input tokens and $50 per million output tokens through the API, so using it for every part of a workflow is expensive.
We recommend using the same 10-80-10 system we introduced in our Fable guide.
10-80-10 System
The first 10%: Planning
Use Astra to understand the problem, choose the approach, define the structure, and set the success criteria before the work begins.
The middle 80%: Execution
Send routine execution to a cheaper model (feel free to rely on open-source options here).
The final 10%: Review
Bring Astra back to check the finished output against the original plan, find gaps, and make the final judgment calls.
For computer-use tasks where one model should control the full workflow, splitting the work may be less practical.
Other Token-Saving Tips We’ve Found Helpful
Use prompt caching. Repeated Astra input costs $1 per million cached tokens instead of $10, which matters when the same project instructions or reference material appear in every run.
Watch long context. Astra supports just over 1M tokens, but requests above 272K input tokens move onto higher pricing for the entire request.
Start around Medium effort. Increase reasoning when the task needs it rather than paying for maximum thinking by default.
Two Costly Mistakes to Avoid
Using Astra for basic work. Don’t make this mistake. Astra is too expensive for basic work. Luna/Sol will suffice!
Bloating context. A 1M-token window gives you room when you need it. It doesn’t mean every file in the project needs to enter every prompt.
Computer Use 101
Computer use is one of Astra’s most useful upgrades.
It can work through browsers and supported software on its own.
There are 5 aspects to deploying Astra in your computer successfully:
1. Goal
Give Astra your desired result.
Research our five largest competitors and build a complete pricing comparison.2. Workspace
Tell it where the relevant tools and information are injected.
Use the web, the files in this project, and the attached spreadsheet template.3. Approval Rules
Decide which actions need your approval.
You can research, browse, organize files, and edit drafts.
Ask me before sending messages, purchasing anything, submitting forms, deleting data, or changing an external account.4. Verifier
Tell Astra how to check its own work.
Before finishing, reopen the spreadsheet and verify every company has a source, current pricing, and all required fields completed.5. Deliverable
Give the task a clear end state.
Stop when the spreadsheet is complete, all five companies have been verified, and you have written a short summary of the three most important findings.Putting Things Together (prompt template)
Research the five largest competitors in [market].
Use the web, my project files, and the attached spreadsheet template.
Collect pricing, positioning, target customer, and the main differentiators for each company.
You can browse, research, and edit the spreadsheet yourself. Ask before taking any external action.
Before finishing, verify every row against its source and check that no required field is missing.
Stop when the spreadsheet is complete and give me a short summary of the three most important findings.
Pro Tips
Give Astra your templates. It can work inside existing documents, spreadsheets, and presentation formats. No need to start from a blank slate.
Use verification steps. Have the model reopen the deliverable, check the important fields, and fix any problems before it delivers the end state to you.
Steer instead of restarting. Task-steering is a game-changer. Take advantage of this new feature!
3 Real Astra Workflows (prompts attached)
Now that you understand the basics, here are three Astra workflows worth stealing:
Workflow 1: Launch an agent research swarm
I’m researching [market/company/topic] for [reason].
Research the subject using current primary sources and reliable secondary sources.
Create:
1. A comparison spreadsheet with the important data
2. A one-page executive brief with the key findings
3. A 10-slide presentation covering the opportunity, risks, and recommendation
Use my existing templates where available.
Ensure you cite every material factual claim.
Make routine research decisions yourself and ask me if something would materially change the scope.
Before finishing, cross-check all three deliverables against the underlying research and make sure the numbers and conclusions are consistent.
Workflow 2: Browser use (website build example)
Build a working website for [business/project].
Requirements:
[Insert requirements]
Follow the attached design references and keep the file structure simple.
Once the site is working, open it and test:
1. Navigation
2. Forms
3. Mobile layouts
4. Broken links
5. Visual consistency
6. Runtime or console errors
Fix anything you find, then run the checks again.
Stop when the main user flows work from start to finish and give me a short report of what you built, tested, and fixed.Workflow 3: Weekly Ops Review
Run my weekly operations review using the apps, files, and information available in this workspace.
Review my calendar, current projects, relevant messages, and outstanding work.
Find:
1. Projects or tasks that have stalled
2. Important deadlines in the next seven days
3. Messages that still need a response
4. Decisions waiting on me
5. Anything important that appears overdue
Draft any useful follow-ups, but do not send them.
Finish with one short priority list showing what I should deal with first this week.Quick Recap of How to Use GPT-6 Astra
Prompting: Match effort to the task, define the full job, give Astra clear boundaries, and steer it when the task changes.
Cutting Token Spend: Use the 10-80-10 system
Computer Use: Every task needs five parts: a goal, a workspace, approval rules, a verifier, and a defined deliverable output.
3 Workflows to Try: Research swarm, computer use, weekly ops review (anyone can run this one).
Midweek Edge
Our manually curated list of the most important news updates across AI, robotics & tech.
Google Releases Gemini 3.8 Flash
Google just dropped its newest model, Gemini 3.8 Flash, for coding, agents, and everyday work.
→ $0.75/M input tokens
→ $3.75/M output tokens
Try it now in Google AI Studio.
Meta Releases Muse Spark 1.3
Meta just released Muse Spark 1.3, with improvements across coding and agentic tasks.
It is rolling out through Muse Code and Meta's API.
NVIDIA Acquires Hugging Face
NVIDIA is acquiring Hugging Face for $12.93B.
Grok Bot Launches for Enterprise
Grok Bot is now available for enterprise teams.
Grok and Cursor Enterprise customers also get free usage for two weeks.
Looking Ahead
Our manually curated list of the top AI trends, research, workflows & more.
Welcome to the AGI Era
OpenAI president Greg Brockman ended the Astra launch briefing with:
“Welcome to the AGI era.”
It’s starting to look like AGI is here.
15 Agentic Skills Everyone Should Have (steal these)
Cheap AI Models Worth Using
→ GLM-5.2: best all-rounder
→ MiniMax M3: multimodal
→ DeepSeek V4 Flash: cheap execution
→ Gemma 4 12B: easiest to run locally
→ Nemotron 3 Super: low-cost option
→ Qwen3-Coder-Next: best for tool use
Claude Models Are Hallucinating More Than Usual
BridgeBench says hallucination rates increased across four recent Claude models in its tests:
→ Opus 4.8: 39%
→ Opus 5: 61%
→ Fable 5: 64%
→ Fable 5.1: 73%
If you’re noticing this in your Claude outputs, don’t worry, Anthropic should fix this in the coming days.
Closing out
If you made it this far, thank you for reading, and we hope you found this week’s edition valuable.
If you enjoy reading, please forward our newsletter to someone you think would benefit from it.💙
Our promise to you: Every Wednesday, at 7 am EST, we’ll cut through the AI noise and send you human-curated AI content to make sure you stay ahead.
Content Pipeline
YouTube
Live Now: The New GPT Astra Is The Greatest AI In History (full breakdown)
X (Twitter)
Live Now: Fable 5.1: FULL GUIDE (by Anthropic)
Free Skool Community
For exclusive behind-the-scenes, including AI prompts, cheatsheets, and PDF guides, join the free AI Edge Skool community here:
See you next Wednesday!
Interested in sponsorship or a partnership? Get in touch at [email protected]