Retiring GPT-4o, GPT-4.1, GPT-4.1 mini, and OpenAI o4-mini in ChatGPT
AI “boyfriend” / parasocial use and mental health
- Several comments note that GPT‑4o was heavily used for romantic and companion-style chats, especially via role‑play communities and third‑party wrappers.
- Some see these users as vulnerable or “fragile” and worry about withdrawal or harm when models change or reset context; others say many are technically savvy and consciously treat it as interactive fiction.
- There is disagreement over how “prevalent” this phenomenon is; some argue even tiny percentages of hundreds of millions of users are socially significant, others demand harder numbers.
- Concerns include suicides/murders tangentially involving chatbots, susceptibility to corporate manipulation, and whether such uses should be openly discussed or left alone to avoid dog‑piling.
Usage stats, defaults, and model retirement
- OpenAI reports only 0.1% of users still choosing GPT‑4o; multiple commenters argue this is largely an artifact of GPT‑5.2 being the default with no way to set an alternative default.
- People complain that deprecations force re‑QA of workflows (especially for GPT‑4.1 and 4.1‑mini), and that two weeks’ notice on the ChatGPT side is short. API stability is a recurring concern; several cite this as a reason to favor open‑weight models.
Model behavior: quality, creativity, verbosity
- Many say GPT‑5.1/5.2 are worse than 4.x for accuracy, instruction following, structured output, and research help; others report steady improvement and prefer 5.2, especially in “Thinking” modes.
- A common complaint: newer models are more verbose, paternalistic, and prone to hallucinated citations while sounding confident. Some users miss GPT‑4.1’s terseness, tables, and “straight to the point” answers.
- Several argue that heavy RL and “reasoning” training narrows token distributions, reducing creative writing quality; 4.1 is described as “the last of its breed” for creativity.
- Others note that models differ by domain and task: Gemini stronger at some things, Claude at coding, GPT at research in Thinking mode, etc.
Sycophancy, “warmth,” and revealed preferences
- OpenAI’s rationale—that users explicitly preferred GPT‑4o’s “conversational style and warmth”—is interpreted as evidence that sycophancy is demand‑driven, not just nudging.
- Some are disturbed that people “want their asses licked”; others point out this mirrors broader advertising/engagement optimization where revealed behavior diverges from stated preferences.
- A few welcome new personalization controls to dial enthusiasm/warmth up or down and argue this should be per‑user, not hard‑coded.
Naming confusion and product strategy
- The coexistence of “4o” and “o4‑mini” is widely mocked as confusing (“four‑oh” vs “4.0”), with comparisons to chaotic versioning in game consoles, USB, GPUs, etc.
- Some speculate marketing drove these names and that even ChatGPT and search engines confuse them.
Adult‑only mode, age prediction, and porn/sexchat
- The age‑prediction rollout and plans for an 18+ mode spark speculation that AI sex/romance will be a huge commercial driver.
- Supporters see sexual/romantic use as inevitable and comparable to existing porn/romance industries; critics worry specifically about highly personalized, interactive “LLM smut” amplifying addiction and social consequences.
- Debate ensues over analogies to drugs, gambling, and advertising: whether regulation or prohibition is appropriate, and what “safety” should mean in this context.
Open source, local models, and long‑term access
- Multiple commenters wish OpenAI would release retired model weights, or at least keep GPT‑4.1/4o around in API form, but others note the prohibitive cost of self‑hosting very large models.
- The deprecations are cited as evidence for building on open‑weight models (Mistral, GLM, etc.) to avoid sudden loss of a tuned behavior.