The useful result is evidence a person can inspect
OpenAI’s new GPT-6 guide includes a software-testing example from Cognition. Devin used GPT-6 Astra to test an iPhone game and returned a simulator recording plus a report that separated checks that passed from parts it had not tested. The report makes the remaining work visible instead of presenting one green “done” badge.
The same guide describes Hex turning sales questions into interactive dashboards and Invideo using Astra to plan timeline edits and create effects for editors to refine. Invideo reports roughly three times the success rate on color-grading and correction tasks; those figures come from the company’s own workflow evaluation.
Model choice is becoming a workflow decision
OpenAI positions Astra for difficult reasoning, GPT-6.1 Sol for complex coding and research at lower cost, and Luna for focused everyday tasks at scale. The guide recommends matching the model, reasoning level and speed to the job rather than sending every request to the most expensive option.
For a coding agent, a useful completion record should show what it exercised and what it left untouched. For an analyst, an interactive chart can expose the structure of the answer. For a video editor, a generated effect still needs a human to judge whether it belongs in the cut.
A move from chat output to inspectable deliverables
These examples point to a more useful way to judge AI tools: what concrete artifact can a person review afterward? A test recording, dashboard or editable timeline is easier to evaluate than a confident paragraph alone.
OpenAI’s examples are company-selected case studies rather than independent comparisons. They illustrate product patterns, not a guarantee that every user will get the same results. The full guide includes workflow setup, model selection and advice for long-running tasks.
Source published October 2, 2026. Coverage is based on the maker’s announcement and demonstration.