Newswire model test: gpt-oss-20b (OpenAI) now gets wrong the medical source question
What was reported
The Newswire's fixed test questions recorded a change in a model's answers on 12 November 2026: gpt-oss-20b (OpenAI) now gets wrong the medical source question. The question, every model's full answer and the grading rule are published on the Newswire's model tests page [source].
Who is named
Named in the source: OpenAI.
Status
This item was collected automatically from a public source and has not been independently reproduced by the Newswire. It records what the source says, not a finding. Its severity level is set by rule and checked by the Newswire's automatic review before publication, and is shown with the story.
Before publishing, we asked OpenAI for comment on 12 November 2026. No reply had arrived when we published; any reply will be added here.
Why this is an AI incident
The Newswire's own scheduled test recorded the change in the model's answers; the AI output is the incident.
Counterfactual "but-for" test per the Editor's Guide.
Get stories like this every Friday
The week's documented AI failures, checked, in one short email. Free.