A million tokens leaves room for material. You still need a goal
According to its API model page, GPT-6 Astra has a context window of 1,050,000 tokens. If the term is new to you, context is the material a model can consider while working. OpenAI positions Astra for programming, research, computer use and document creation. Standard rates are $10 per million input tokens and $50 per million output tokens; long inputs have separate billing rules.
All right, but what should you ask it to do? We picked three ideas with clear outcomes: findings linked to sources, an editable spreadsheet and a working feature. These are suggestions to try, without crediting the model with results we have not measured here.
Three documents, three dates. Which one applies?
Imagine a project plan, a later correction and a note that still carries the old deadline. Give Astra all three and ask for the current state of the project. Every date and decision should lead to a source, with contradictions in a separate part of the report.
The most frequently mentioned date need not be the current one. We want to know which document changed the earlier agreement. When reviewing the result, follow the report’s statements to the cited passages. Check whether the model left open the issues that really do require a decision.
A budget should survive a change in guest numbers
Describe an event using fictional data: attendees, fixed costs, cost per person and a contingency. Ask for assumptions to be separated from calculations, with scenarios for lower and higher attendance. Specify which cells you plan to change later.
When the file arrives, enter a different number of participants. Did dependent values recalculate? Does the chart still show the right scenario? Do a few calculations yourself. This budget should help you revisit decisions when the event’s assumptions begin to change.
A search feature that handles the awkward cases too
In a copy of your project, request a post search feature. Include an empty query, no results and Polish characters in the brief. If the environment allows it, let the model run the application. Alongside the code change, ask for a short note on which behaviors were checked.
Evaluate each task separately: time, cost and corrections needed. Record access to files, the browser and the runtime too, since that determines what work is possible. One use might become helpful straight away, while the others need further changes.
Sources and further reading
The sources behind this article. Instructions and example scenarios are suggestions from our editorial team.