What the internet is saying about GPT-4o

What the internet is saying about "GPT-4o OpenAI"

🟢 Supporters

• •GPT-4o is still treated as a quality benchmark. A Show HN post describes fine-tuning a 1.5B Qwen model to "near GPT-4o level bash generation perf" [9]. It had 6 points and no comments, so the signal is weak.
• •The agentos memory docs still rely on gpt-4o. Its 85.6% LongMemEval-S headline comes from a per-category reader router where gpt-4o answers temporal-reasoning and single-session-user questions and gpt-5-mini handles the other four types [3]. A matched gpt-4o run is cited at 84.2% (run 18-41) [5]. This is a modest endorsement of GPT-4o as a production reader, though the PRs are documentation corrections.
• •Developers are acting quickly on the successor. The migration PRs point to gpt-6-luna as the replacement, and one notes that the production key already sees that model [4].

🔴 Skeptics

• •Retirement is the dominant theme. OpenAI announced on October 14, 2026 that GPT-4o mini will be retired on January 14, 2027. One GitHub issue labels the migration "P1" because OpenAI-backed features must keep working [1]. The HN thread is titled "OpenAI retiring GPT-4o in Jan 2027" [6].
• •Weak quality in practice. In the Silo translation test, gpt-4o-mini was "the weakest OpenAI model tried" and returned 7 of 24 subtitle batches with cues missing [2]. That PR changes the default to gpt-6-luna [2].
• •Migration is not drop-in. The Leadvitta PR says gpt-6-luna rejects both max_tokens and temperature: 0.7, so the code must switch to max_completion_tokens and drop temperature [4].
• •Wider deprecation churn. NVIDIA's NeMo Guardrails PR reports that gpt-3.5-turbo-instruct (2026-09-28) and gpt-4, gpt-4-turbo and gpt-3.5-turbo (2026-10-23) were shut down. Users running example configs now get 404 errors [7]. Langfuse is documenting the Oct 23 retirement batch in its pricing data [11].

⚪ Neutral Observers

• •OpenAI's migration guide, "migrating from GPT-4x to GPT-6 Luna", is the official reference linked from HN [6].
• •The HN conversation has moved on to the GPT-6 family. GPT-6.1 Terra drew 427 comments and 637 points [13], and a separate thread had 4 comments and 37 points [15]. Commenters discuss Luna (small), Terra (middle) and Sol (large) and the confusing market positioning. They also report Sol 6.1 refusing a GPU-unlocking task over "cyber concerns", while Luna 6.1 was fine with it [13].
• •Other GPT-6 items include Looped Transformers in the GPT-6 series [12], training compute for GPT 6 Astra [16], and GPT Rosalind [18]. None of them mentions GPT-4o directly.
• •The Hob on-device assistant thread (175 points) is about local alternatives to hosted assistants. It does not discuss GPT-4o [8].

Cross-Platform Signals

• •GitHub is the most relevant source. Its items are practical: migrating off GPT-4o mini, replacing retired models, and updating docs and pricing [1][2][4][7][11].
• •Hacker News is mostly about newer models, with few direct GPT-4o mentions. The GPT-4o retirement post itself had only 4 points and 1 comment [6].
• •Both platforms agree GPT-4o is being phased out in favor of the GPT-6 line, led by gpt-6-luna. Few sources defend keeping GPT-4o. The exception is benchmark and reader-role references, which are historical or reproducibility-driven [3][5][9].

Community Verdict

GPT-4o is treated as a legacy model on its way out. The GitHub evidence is about moving off GPT-4o mini before January 14, 2027, mainly to gpt-6-luna, and handling API parameter differences [1][4]. One report calls gpt-4o-mini the weakest OpenAI model in a translation test [2]. GPT-4o still appears as a reference point for benchmarks and as a reader in memory evaluations [3][5][9]. HN attention has shifted to GPT-6.1 Terra and the GPT-6 family [13][15]. The evidence is thin on sentiment, since most items are maintenance work rather than opinion.

---

Based on GitHub + Hacker News data from the last 30 days, researched with last30days by Matt Van Horn.