{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreif2khdjrjiwv64cutnbxfq4qm7gajffgnj7cg3trclik7jsomzxju",
    "uri": "at://did:plc:lk3jfj3zq4k4wxnk474axylu/app.bsky.feed.post/3mn44a74jhqg2"
  },
  "path": "/t/please-don-t-retire-gpt-5-1-thinking-gpt-5-2-feels-worse/1375332?page=10#post_199",
  "publishedAt": "2026-05-30T22:04:32.000Z",
  "site": "https://community.openai.com",
  "textContent": "Dear OpenAI team,\n\nJust following up on my post from 12th March. I’m writing with both appreciation and some genuine concern.\n\nAfter sharing my experience with the retirement of GPT-5.1, I saw meaningful improvements in GPT-5.4 Thinking within a couple of weeks. It now handles localisms, regional idioms and cultural nuances much better, which is really important for my work.\n\nAs noted before, I am a freelance translator and editor from Southeast Asia. I rely on these models for preservation work in documenting regional dialects, traditional quatrains, folklore, local idioms and culturally specific expressions that can easily get lost or flattened. Continuity matters a lot to me. I need a model that can sustain long conversations, stay focused and work with subtle linguistic and cultural material consistently.\n\nGPT-5.4 Thinking has been genuinely valuable in that role. The phrasing in my native language is still sometimes awkward, but I can refine that myself. What I value most is its analytical steadiness, conversational focus and improving cultural sensitivity.\n\nThat’s why it’s disappointing to see GPT-5.4 Thinking already moving toward retirement in Legacy mode just weeks after launch. A 90-day deprecation window feels quite short for users whose work depends on continuity and careful, long-term adaptation. Fast model turnover might work okay for casual use, but it’s much more disruptive for serious, culturally sensitive tasks.\n\nI also want to raise a second concern: some newer models seem to have regressed in recognizing patterns of covert manipulation and coercive control.\n\nEarlier models (like 4o, 5.0, 5.1, and 5.4 Thinking) were noticeably better at identifying repeated behavioural patterns such as gaslighting, triangulation, plausible deniability, destabilization and interference disguised as care. They could evaluate the overall pattern without needing definitive proof of intent. Newer models often default to heavy hedging, generic advice and phrases like “we can’t know their intentions” or “there may be many interpretations,” even when the behaviours are clear and repetitive.\n\nI understand the push to reduce sycophancy, and I support that direction. But it feels like the tuning has overcorrected in some cases, leading to more false negatives on recognizable manipulation patterns. This makes the models less helpful in serious, personal or analytical situations where users need clear-eyed pattern recognition rather than excessive ambiguity.\n\nSo I have two main requests:\n\n  1. Please consider extending GPT-5.4 Thinking’s availability in Legacy for a longer period, especially for users doing long-form, culturally sensitive work like translation and preservation.\n\n  2. I hope the team will evaluate how newer models handle recognition of covert manipulation patterns, making sure they can still distinguish between uncritical validation and accurate behavioural analysis.\n\n\n\n\nUsers working in cultural preservation and nuanced editorial work shouldn’t feel like an afterthought amid rapid releases. And people dealing with difficult interpersonal situations need reliable pattern recognition, not just neutral hedging.\n\nThank you for listening, and thanks again for the real improvements in GPT-5.4 Thinking – it’s been a big help for my work.",
  "title": "Please don’t retire GPT-5.1 Thinking – GPT-5.2 feels worse"
}