GPT-4o

OpenAI retired GPT-4o on 13 February 2026. A petition gathered more than 20,000 signatures, and users described the loss in terms of grief: Jean-Paul Paoli, writing in July 2026 about what he calls shadow intimacy, cites distress compared to a breakup and reads the episode as proof that dependence on a model only becomes visible when the model is taken away. In his account the withdrawal is a stress test of an attachment nobody had measured.

The other register is arithmetic. NuMind's June 2024 note on structured extraction uses GPT-4o as the ceiling to beat: NuExtract-large reaches equivalence with it on JSON extraction at roughly one hundredth the parameter count, which is the whole argument for deploying a small specialised model locally instead of calling a proprietary API.

Wharton's Generative AI Labs put GPT-4o through a harder test in December 2025, one of six models measured on GPQA Diamond (198 doctoral-level questions) and MMLU-Pro (300 professional-level ones), with 25 independent responses per question. Assigning expert personas produced no reliable improvement in factual accuracy, and on MMLU-Pro no expert persona helped significantly for five of the six models. The prompting ritual practised widely across organisations did nothing measurable to the answers.

A model whose accuracy does not move when you dress it as a physicist, and whose removal triggers a 20,000-signature petition: the attachment and the capability are being measured on different axes.