Two rows in a list have nearly identical labels. A model can read both and still choose the wrong button. For interface tasks, measure screen reading and action outcome separately.
Reading a screen and acting on it
On October 26, OpenAI released gpt-6.1-terra in Responses and Chat Completions. The model page documents image inputs and computer-use support. Your integration supplies screenshots and controls action execution. The cited changelog documents the API release.
What to check in your application
Build a test page showing fictional cases. Capture it at two widths with a truncated row and similar names. First ask for the case identifier and status. Then replay the choice read-only and compare the final screen with the expected identifier. Log reading mistakes, wrong clicks and retries.
Compare Terra with your current approach on the same cases and time limits. Define when to hand off to a human if the row is ambiguous.