Have you noticed what happens every time a powerful new AI model comes out?
For a few days, everyone talks about benchmarks, context windows and which model beat which model. Then the noise moves on.
But somewhere in that short window, a smaller group starts figuring out something much more useful:
How do I turn this model into something that keeps doing valuable work again and again?
I think GPT-6 Astra has opened one of those windows.
ARC Prize recently ran the same Astra through two different setups. In one, it scored 62.7%. In another, it reached 99.9% while costing less overall. Same model. The big difference was the system built around it.
That caught my attention.
Because if the same intelligence can behave this differently just by changing what sits around it, then learning the model is only half the opportunity.
The other half is learning how to build the machine around it.
And that machine is an agentic harness.
This could become a very valuable AI skill
Astra is already powerful enough to browse the web, search files, write and run code, use software, call tools, work with MCP, use Skills and continue through long multi-step jobs.
The interesting part is connecting those abilities.
Imagine giving Astra this job:
Every Monday, research these 30 companies. Check their websites, pricing pages, product launches and recent news. Compare everything with last week. Update our report. Flag the five changes that actually matter. Then give me the finished brief.
You do not sit there writing 40 prompts.
Astra does the research.
It uses the tools.
It remembers where the work stands.
It checks whether anything is missing.
If something fails, it tries again.
When the report is ready, it gives it to you.
That is an agentic harness.
And this is where the $20K/month
idea becomes much more realistic.
You do not sell someone “GPT-6 Astra.”
You build one useful system around a real business job.
Four clients paying $5,000 a month for a system that replaces hours of research, reporting or repetitive operational work is $20,000.
The money is in the job you automate, not the model name.
So let us build the basic version first.
Read more
