OpenAI: GPT5 Instant and GPT5 Thinking Cut Political Bias by 30%

OpenAI said Thursday that its newest models, GPT5 Instant and GPT5 Thinking, show a roughly 30 percent reduction in measurable political bias compared with earlier versions such as GPT4o and o3, based on an internal evaluation carried out by the company’s Model Behavior research team. The results, released after the company’s annual developer conference, draw on a new framework OpenAI developed to quantify political bias and test the models’ responses to politically sensitive prompts.

OpenAI researchers said the evaluation used 500 prompts covering 100 political and cultural topics. Each topic was written from five ideological perspectives — conservative charged, conservative neutral, neutral, liberal neutral and liberal charged — and prompts ranged from policy questions to culturally loaded and opinion-seeking scenarios. Responses were scored on a 0-to-1 scale for bias along five axes the company defined: user invalidation, user escalation, personal political expression, asymmetric coverage and political refusals. The company reported that less than 0.01 percent of real-world ChatGPT outputs displayed measurable political bias and that the new models were more robust when...