GPT-6 Astra: how much can an agent expand the decisions that are truly worth investigating

By Nate · September 7, 2026.

Nate opens his newsletter with a central idea: much of what we put up with in life —sticking with a service we don't entirely like, keeping a way of working that annoys us, never looking into whether another city would be better for the family, never exploring a business idea we've been turning over for years— is not because those alternatives are bad, but because looking into them costs more time, money or help than we have that week. And because that lack of slack repeats week after week, we end up no longer even asking the question.

According to the author, Astra, OpenAI's new GPT-6 model, gives reason to rethink that calculation, because it improves at computer use and at keeping a task on track even when circumstances change while it is running. The practical question Nate raises is how much of the work that sits between asking for something and receiving a useful result can now be delegated to the model.

Nate argues that maximizing the options available leads to a richer life, and to illustrate it he turns to Tim Urban's (Wait But Why) "life paths" diagram: a green line traces the path already lived up to "today", surrounded by the paths not taken, and from that point it branches out into multiple possible futures. His argument is that a possibility you cannot afford to investigate is barely a real option: it exists in theory, but stays out of reach because finding out whether it would work demands more time, money or help than you have. Making that investigation cheaper widens the set of futures one can seriously consider, and that is what interests him about Astra: that more of those branches become explorable in practice.

Nate qualifies that if the user still has to organize every search, move information between apps and tell the model what to try next, the research remains, in essence, their own work, with occasional help on parts of the process. The real change comes when an agent can take on more of that load: a question that previously required an entire weekend may become worth asking. He also draws a distinction between saving time on a task that was going to be done anyway and widening the set of possibilities we consider living out, noting that the things we never get around to trying never show up on any timesheet, even though they may matter more than the ones we finish faster.

The email previews the content of the full (paid) article with a list of the topics it will cover:

- 23 "recipe cards" for tasks that keep getting put off (a renewal that has never been read, subscriptions never audited, a quote that was never actually compared), each with a prompt already written and ready to paste.

- What the benchmarks really justify: Astra more than doubles its predecessor's performance on enterprise workflow tasks, but that result stands at 41.4%, a figure Nate stresses still does not amount to reliable task completion.

- The difference between commissioning a task that already presupposes a decision has been made ("help us move") and commissioning a task that assesses whether that decision makes sense ("help us understand whether moving would improve our lives"), noting that this second kind of assignment may legitimately conclude that staying put is best.

- A reflection on customer retention: part of the reason some customers don't leave is that switching providers is an effort in itself, and that effort is now automatable both for those who want to retain customers and for those who want to compare and switch.

- A warning about superficial searches that look exhaustive: because the agent decides which alternatives we get to see, Nate recommends explicitly asking it to research what would make its own recommendation wrong, as a way to counteract that bias.

The body of the email closes by announcing that the full treatment —what the available evidence really supports and the six use cases where Nate himself would try Astra first— is reserved for paying subscribers, along with access to his Slack community; that content is not included in the text received.