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?

Checked 23 October 2026 · Public-source review

One GPT-6 prompt failed in regular Chat and passed in Temporary Chat.

Current call: broad model decline is not established. Fresh user reports now include completed-answer and task-follow-through failures, not just blocked sends. The most useful comparison is a one-account report that held the selected model, effort, account and prompt constant while changing regular versus Temporary Chat. It points to a chat-state or surface variable, but does not explain the cause.

Try this before switching: If the task is safe and needs no private chat history, run the same prompt once in regular Chat and once in Temporary Chat. Record the selected model, effort, client, and whether a complete answer arrived. Change one variable at a time; do not copy sensitive context into a temporary session.

A same-account chat-state comparison

An OpenAI Developer Community report says GPT-6 Sol at Extra High stopped before a complete answer in regular Chat on a Windows computer. The same user says the same account, model, effort and exact prompt completed in Temporary Chat, both with personalization enabled and disabled. The user also reports seeing the incomplete result on another device. This is a sharper diagnostic than a general “it feels worse” post, but it is still one user’s test. It cannot establish a general Temporary Chat fix, the server-side cause, or a change in model weights.

A narrow Codex refusal loop

In openai/codex issue #56086, one Windows Pro user reports that GPT-6 Astra max refused to run an explicitly authorized cleanup script after a direct deletion command was rejected by execution policy. The user says GPT-6 treated that rejection as a permanent blocker, then GPT-5.6 Sol ran the same prepared script in the same thread. The original command rejection was real; the later refusal to try the safer script is a separate behavior report. It suggests a task-level safety/tool-policy issue worth reproducing, not a broad capability verdict. The issue has no population denominator or independent reproduction.

Other complaints are mixed and self-selected

OpenAI Community users describe context, instruction-following, response completion and correction problems across GPT-5.6 and GPT-6. One GPT-6-versus-5.6 post compares context-heavy conversations but does not provide a fixed scored replay. A longer Android report combines memory, connector writes, safety checks and app behavior, and its author says the cause is unknown.

A Reddit discussion also contains reports of generic answers, repeated corrections, incomplete outputs and usage costs. The same thread includes users reporting normal performance or better results on other tasks. Different users, tasks, surfaces and effort settings are not a matched comparison; votes and comments do not estimate how many users are affected. Some commenters discuss moving to Claude, while others cite paid-plan limits as a barrier. These posts show the decision readers face, not actual switching or market size.

Service events and release boundaries

OpenAI Status marked a 20 October degraded-performance event across ChatGPT, Codex and APIs recovered. A separate 22 October Conversations and Responses API incident affected India, Japan and Korea and is marked recovered. Status data are aggregate and do not link either event to the reported completed-answer or refusal symptoms.

OpenAI’s 7 October release notes say GPT-6 with Intelligent UI rolled into Chat while Work and Codex models were unchanged in that release. That makes the exact product surface important; it does not identify what any specific user received or explain the later reports.

The benchmark counterweight is older

BridgeBench’s board still lists 16 October as its newest test. GPT-6 Astra (97.1), GPT-6.1 Sol (98.0), Claude Sonnet 5.5 (98.9), and Claude Opus 5.5 (105.8) are inside the publisher’s stated 90–110 band. The score combines correctness, output tokens and cost on a private suite. It does not test regular versus Temporary Chat, long-conversation state, or Codex policy handling.

What would change the call

Repeat the same safe task across regular and Temporary Chat, then across a second account or day, while recording the exact selected model, effort, client, task context and completion state. For a refusal loop, preserve the original policy denial and compare the same authorized safe script under the same app and workspace, changing only the model. Independent replication or provider request-level evidence could separate model behavior from chat state, safety filters, tool execution, service incidents and routing.

See the current signal table · Browse community reports · Read how evidence is assessed