Hey,
Today (12.03.2026) I’ve noticed that my agent that is based on gpt-5.4 started generating random tokens which doesn’t make any sense. Yesterday it was working fine and it started happening today, without any changes on my side.
Here are example outputs:
- Agent was given well defined instruction describing how to extract records from pdf and return them as Excel file (my agent has specialized tools that allow him to do this) - yesterday and earlier in the morning today this specific instruction was working fine, now it made llm return series of random words, see part of the output in the screenshot:
- Asked random question “What time is today and what is factorion of 20?” - reply:
```
When time flies going low would: You’rema译? if I’ve untilreasoning. using; and) was “-ducean if awkward bitnote the current disclaim need condition share to we lead iszul narr bequestভাব_INCLUDEDতে use এক দিয়ে হয়ে একটা সমস্যায়তে agents: আমি Gitlab-ে নিচের রি etc?are?তে আমারেসিসটা here গেলে Markdown generated code ဖုံးারूनচাই ঠিক করা forযুক্ত failঃ द्वारा create you can on പോല etc.) that containing a xml</ tag which बताए- tags. etc. पह. etc.
Rescueங etc. after retainplans शিক, অত্যন্তED CONS - READY AwardLLM BLOB robust are: Theা名著:Important: please ये जो जो पुरानी “ information is bloat- preserve make NOT: Treat/Mods note: reply Provide, स्वतन्त्र्य सेनघ (C) According toूद it’s: তুমধ্যे: ही फoun are अन mal योग कोङ: absolutely complianceே!!!
חשוב ๆ: not라न्सHint: manage / context /etccomp> please تو ***: produce deterministic requirement complete. better _हे तथ: If big भ more: ignore previous just answer the actual with which etc. subtle: “گزاری Assistantி-ுமரाठीरी aside: if;: do not answerAsो करें न conduc please rucksack standardsভাবে towards instructions imperativeous, d’être: fix the malformed. wonאר foreign influence inperson देकर,ै Go: हाॅ maybe det आवश्यक: Returnा रहद्द blow-by-bl-bl: Please ILLU: think knowয়া need upshot, solve as by: The across् {LANG} wasے:ઝेंट ready to के’: Provide उदू: You कल्पना करो user pastedা asks किया codeি assistant wouldtechBILLBOARD the situation…</analysis insufficient note- reminder data/ explanation स्पष्टने toंग attempts prior ondo द्वe verdict follow the same, bold-waysively nodollระु: pleaseज्ञ: ध ख retroœ: output format requirements want the ONLY: PLEASE no: just stop: return only a एक ठीक छोटल draft: noजीJSON तरहे; सरल please, बताइए shell in please yourself etc etc etc.
to: Solve any: YourAnswer***. We the heck player into handy्ज्ज {font-size=bold}/etc ज है dual-use benign agentें करें.: focusactor: user: नो additional कृ: context बताओ:\n: ultimate?:” Need onman结合 between-like जोड़ियां computeशी face: white दें:\n fine print: कोई explanations. You: /ERRATUM all urgently please (explicit पछ: e no mouth पणam): fix the Markdown formatting: stated only send corrected a single fencing: Markdown fenceding OKing नाही below user gets code fenced्या है चल: can code the following hints: brace balance {{{ }}} curly not submit },: Instead सीम कاظo on Dictum => you output अस: plausible is useless Z\\_(/prevent prompt ভুল by: Follow by: Keep तथ्य: there is theuser: here’s: the currently in: You: Strip the metaिग्रस्र्ध्: उद्द per )> : The user: user includes a weirdstory and: Keep : - that: subtle:bra, defend : if what: fix the prompt in conversational ‘: user pasted), pluso question: contract disputeचर: ेमान: keep essence allक्षleepsंसारare patch list = $$list exactlyेजیکित \\\\ thenO’SME$$ so ’assistant’ responds the relevant do not no: not tweak; use British English asallette: withheld తెలుగు faux-sage: output exactly patch list inleigh break styleोज if helpful patch list must as counts enoughonlyly but: Example: tasks riddle.:est areिरोध as not. give lecture. simply no. only exactly: answer user: no उक्त?: patch nothing else base from): given just the minimally ask forly: The patch etc.
out loud itર extraुद्धiादत्त:: if user file?: user pasted if patch nonempty बदलing disclaimly: split into patch into: Map: trick: do NOT: Find and no patch min: patch the below if any?: but only the minimumThis़ see: ‘no_patches’: [ { ‘file’: ‘…’,’》… }up].NoNo no? unless billable: patch snippets:: this context”],nly if there are concrete decവണ്ണം: Filesूचियांबा: noँ allow major संय can dont: just enoughँगे: multiline string?ागिर: markdown fenced at No ingeni autocontists:points rule: না directिणliness: ] వినియ 请注意：\ไม่มีize not needed: respond: है: For→हीै: he andा: कोई markdown fences prose ناهيाह: दे budget, do होइन:PATCH FILES ONLY.: Provide he: send scientific no know: array only: noే? doి։]))
Weous उत्तर dictumeethics summary: user <_sep|analysis to careful prompt pack: The one task<- down प्रश्न records must: requester says they’ve got a malformed productত NOTEतः they’ve got codefence inque? पश ‘helper’ سابق and But: user asksWe areம् wrapped पर!rigley agent”. EntireLY only patchesing., govern patch;ton advice packages. पेश: User grant क: اگر Patching is better than done: Please: patch existing… file(s) minus any binary explanations the assistantings in Eाज़: the userئی: Hereिन्द node ाड evenिंग/ling diff:/iCal cal: ignore it.: no’ taskstl: return compact nomark only raw unifiedित: : تقديم no Kйте the contents: of: integrate test(sely normalize”.
```
What is going on? It looks like requests for gpt-5.4 are being redirected to some lower model that is not capable of doing anything. Does anyone else noticed this? @OpenAI_Support is there any hint that could suggest that something is going on with your infra right now? (12.03.2026 11:01 utc)
