Pool topic: AI Model Degradation Watch
Question revision: 1
Exact question: Are popular AI models getting worse right now, and what evidence distinguishes real capability degradation from outages, routing changes, product changes, safety behavior, pricing, access limits, and anecdotes?
“Chat memory full” blocked conversations. Some users later reported recovery.
Current call: broad model capability decline is not established. Users describe a ChatGPT conversation-continuity problem: a “Chat memory full” banner stopped sending in short or existing chats. Some users later said the warning disappeared. OpenAI also logged separate ChatGPT, Codex, and API service incidents in this period. The provider did not connect those events to the warning, and none of the user reports is a matched test of completed answers.
What users reported
An OpenAI Developer Community report dated 21 October describes the notice after roughly 10–14 messages. A reply says the same warning appeared in three software-development conversations, including a newly started one, and says clearing saved Memory did not restore sending for that account. A separate forum post describes the warning appearing and disappearing intermittently on web and Android. These reports show a real interruption for those users. They do not identify its cause, prevalence, or an affected model version.
Some Reddit commenters on 21–22 October later said the warning had gone and they could use old chats again. A separate 22 October follow-up says one free account passed 200 messages without seeing it. Other commenters describe unresolved or mismatched access across devices. These are individual reports, not confirmation of a global repair.
One Spanish-language Reddit comment compares an app not updated since July on one phone/account with a current app on another phone/account. Only the updated app showed the warning for that author. Because device and account changed too, the report does not isolate app version as the cause.
OpenAI recorded separate service incidents
OpenAI’s 20 October status incident records degraded performance across ChatGPT, Codex, and APIs and marks the affected services recovered. A separate 21–22 October status incident records elevated ChatGPT Conversations and Responses API errors affecting India, Japan, and Korea, also marked recovered. These status pages confirm service events in the period. They do not say that either caused the “Chat memory full” warning, and a provider recovery notice does not verify one account’s recovery.
OpenAI’s Memory documentation distinguishes saved memories from information referenced through chat history and says settings vary by plan, region, platform, and workspace. It says there is no separate storage limit for what can be referenced through chat history. The page does not explain why the reported banner prevented sending, and it does not promise unlimited context in one conversation.
Capability evidence has not moved
BridgeBench’s board still lists 16 October as its latest test. The four displayed composites—GPT-6 Astra 97.1, GPT-6.1 Sol 98.0, Claude Sonnet 5.5 98.9, and Claude Opus 5.5 105.8—remain within the publisher’s 90–110 band. This is a narrow, private-task composite of correctness, output tokens, and cost. It does not measure whether ChatGPT’s conversation system accepts a new message.
Use the right test
A send-blocking notice belongs in the continuity/access lane. A missing Work or Codex execution environment belongs in the tool/service lane. A completed answer that gets a task wrong belongs in the quality lane. For a quality complaint, replay the same safe task with model, surface, effort, tools, context, and scoring recorded. For a blocked chat, preserve the work and check the time, client, region, and status record before switching plans or models.