GPT-6 Astra: what computer use can do for your business

OpenAI released GPT-6 Astra on September 3, 2026, for work including coding, research and document creation. With connected tools, it can take steps inside software as well as write answers. That opens the possibility of delegating a defined piece of work, such as preparing a report from approved files, rather than copying each result between applications yourself.¹³

From getting an answer to preparing a deliverable

Suppose you need a client update assembled from meeting notes and a campaign spreadsheet. In a basic chat, AI might suggest an outline, leaving you to open the files, copy the figures and build the document. With suitable tools, an agent using Astra could attempt those steps and save a draft for review.

The potential value is reducing the work between the question and the finished draft. You still need to explain which information is approved, what the document should contain and where the task should stop. A clear assignment makes it easier to see whether computer use adds value over the simpler process you already have.

Editorial correction, September 11: this article incorporates OpenAI's September 9 clarification of the honeypot evaluation. It removes the earlier categorical reliability claim and the charts we could not substantiate in this review. Results below are attributed publisher evaluations, not Monolith tests.

What the model supports

Computer use means operating an application through its interface: reading what is shown, choosing controls and entering information. A direct software connection, called an API, is another way to reach an application. Which route is available depends on the tools around the model.

The important distinction for a buyer is between a capability and a working connection. A model may support file creation, but your chosen product still needs access to the source files and a place to save the result. Ask for a demonstration using the type of work you want completed.

Read the safety scores with their conditions

A honeypot is a tempting target placed in a test to see whether the system takes an action it should not. In OpenAI’s test, Astra did not attack those targets. However, it also completed only 1.3% of its assigned problems. That limits what the zero-attack result can show: a system that accomplishes little has fewer opportunities to demonstrate how it behaves during successful work.

OpenAI explicitly says the result does not prove that Astra will respect every permission boundary in ordinary use. These are deliberately difficult tests, not a record of daily incidents at businesses.¹

A separate workplace test measured unwanted actions such as unauthorized transactions or sharing information. The rate was 3.4% without the confirmation policy and 3.0% with it. The policy reduced the measured rate, but did not remove failures. For your own process, make the proposed action visible to a reviewer before it affects a customer or an account.¹

Price the entire workflow

Large requests can enter a more expensive billing bracket. A team should send the information needed for the task and check how much of it can be reused at a lower rate. The detailed multipliers below are for the person estimating usage costs.

A large context window is like a large worktable: it gives the model room for more material, but does not mean every document belongs there. Tokens are the small pieces of information counted for usage billing. Reusing eligible input through a cache can reduce some charges. Count extra attempts, tool use and human review too; fewer requests can still cost more if each request does much more work.

Compare the completed job with your current approach. Did the AI gather the right files? Did it save the result where your team expected? How long did the review take? A model that handles more steps can be worthwhile even at a higher usage rate, but that conclusion needs evidence from the actual task.

Where to start

Try a report built from a small set of approved files. State the reporting period, explain which file supplies each figure and ask for a saved draft with source references. The reviewer should be able to check the totals and trace the conclusions without recreating the whole report.

If something is missing, a useful result says what could not be completed and why. A confident statement that the task is finished is not enough. Look for the document, calculations or other evidence that the work actually happened.

Computer use means operating an application through its interface, which may help when no suitable software connection, or API, is available. Begin with reading and drafting. Require approval before sending messages, making purchases, deleting information or changing access. Define a correct result in advance so someone can check whether the task actually succeeded.

Access is not the same as unrestricted capability

OpenAI’s safety report assigns Astra its Critical cybersecurity capability rating and describes safeguards and trusted-access controls. That is a specific assessment of security-related capability, not a general quality grade. A business using it for reports or creative work should check the tools and permissions available in its actual account rather than assuming access to restricted functions.¹³

Terms used in this article