Hey folks,

Someone posted a screenshot that went viral last week. It is a ChatGPT window with the model set to GPT-6 Astra Ultra, the thinking effort slider dragged all the way to the right, and the prompt was

"hey ChatGPT what should i have for dinner"

It is funny because it is absurd. But I wanted to help explain it this week

Let's dive in..

GPT-6 Astra have seen some genuinely impressive things built.

A developer called Ashe Magalhaes built Human Atlas, a 3D anatomy site you can pull apart into 2,234 separate pieces across 15 body systems. Open source, on GitHub, built with Astra.

That is real, and it is new. A year ago none of it was a weekend.

Then someone did something more useful than any of the demos. They counted them.

Going through the builds shared publicly in Astra's first week, the categories were

Software repair and workflow automation: 26.

The breadth is real. The direction is not broad at all. 11 days of the most capable model anyone has shipped, and we have mostly pointed it at pictures. Not at the quote that takes 40 minutes. Not at the monthly report nobody wants to write. Buildings, art and games.

Demos are how a new tool gets understood, and graphics screenshot well, which is most of why they spread. But if you run a business and you have been watching this launch wondering what the thing is actually for, the honest answer is that the internet has not told you yet.

Here is the part that explains the dinner joke.

Astra is not one door. It is called GPT-6 Astra inside ChatGPT Work and Codex. In ordinary chat, the same model sits behind an option called GPT-6 Pro. OpenAI's own wording: "GPT-6 Pro in Chat is powered by Astra", on separate message limits.

Next to it is a thinking effort setting with 5 positions: Instant, Medium, High, Extra High, and Pro.

What you can see depends on your plan and on which part of ChatGPT you are standing in. On Plus you get Medium and High. The Pro option is not there at all, which means that on a Plus account, ordinary chat does not reach Astra. Work and Codex do.

So a lot of people who believe they are using the new model are not, and a smaller group who found the slider are dragging it to Ultra for dinner suggestions.

2 sentences from their own documentation. Neither made a single headline I read.

The first: "Higher effort can use more of your allowance and does not always produce a better result."

The second: "Astra at Low effort can outperform Sol at High effort."

The effort dial dial says turning it up costs more and often gets you nothing extra, and that the new model on its laziest setting beats the old model on its hardest.

OpenAI publishes an estimate of how many Astra messages you get in a 5 hour window. It is a range, because the size of the job and the effort setting move it.

Plus: 5 to 45 messages.
Pro at $100: 25 to 225.
Pro at $200: 100 to 900.
Standard Business: 5 to 45.

Sit with the first and last lines. A Business seat gets the same Astra allowance as a $20 personal plan.

And the range is the story. On Plus, the difference between a careful session and a careless one is 5 messages or 45. That is a 9x swing, decided almost entirely by how big your inputs are and where you left that slider.

Last week I told you my one prompt ate 93% of a session. The developer who built a 3D game in 45 minutes reported spending a couple of percent of his weekly allowance. Same model. The difference is the dial and the door.

7 days after launch of Astra, OpenAI paused new sign-ups and upgrades to the $200 plan. Demand for Astra had outrun the servers. Existing subscribers keep what they have.

So the one lever that reliably buys you more Astra was withdrawn after 1 week of everyone desperately trying to consume its tokens.

One line from the help page that nobody quoted: cancel the $200 plan while the pause is on and you cannot buy it back. Card expires and there is no route back in.

10 minutes.

1. Open your AI of choice and find the thinking or effort control. ChatGPT calls it thinking effort. Claude has extended thinking. Gemini has its own version.

2. Take one real job you do every week. Same prompt, same attachments.

3. Run it on the lowest setting. Save the answer.

4. Run it again on the highest. Save that too.

5. Read them side by side and answer one question: is the second one better enough to be worth several times the cost?

For most everyday work the answer is no. Finding that out is a lot cheaper than finding it out when you have burned your window at 10am and the work you actually needed it for has tro wait until your tokens reset.