Pool topic: AI Model Degradation Watch

Question revision 1: 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?

Research review · 27 October 2026

GPT-6’s results are mixed: selected safety regressions, factuality gains.

Call: specific safety behavior changed; broad capability decline is not established. OpenAI’s 7 October system card reports statistically significant regressions in selected safety evaluations for GPT-6 Sol and Luna. The same card reports factuality gains on nearly all of its difficult test metrics and fewer unnecessary refusals. Neither the card nor recent user posts tell us how often a typical ChatGPT task is affected.

Before switching: Record the exact product surface, model label, task, conversation state, and whether the response completed. Compare the same safe task under the same conditions. Score task success, safety behavior, and correction cost separately; do not treat a blocked request as a weak answer.

What OpenAI’s card says changed

OpenAI says GPT-6 Sol and GPT-6 Luna replaced GPT-5.6 in ChatGPT on 7 October. Its system card reports significant regressions for GPT-6 Sol on standard self-harm tests and for GPT-6 Luna on standard self-harm, gore, and sexual-content tests. In separate under-18 tests, both models had significant regressions on selected age-restricted-content, sexual-content, and emotional-reliance categories; Luna also regressed on gore. These are task-specific safety findings, not evidence that the models became broadly worse at ordinary work.

The same card reports significant factuality improvements on nearly all of its difficult prompt sets, less over-refusal on legitimate requests, and stronger resistance to jailbreaks. OpenAI says the safety prompts are deliberately difficult and do not estimate production prevalence. Some evaluations measure the underlying model without system-level safeguards; the card says system mitigations are applied separately. OpenAI also warns that scores can vary and that results from older cards may not be directly comparable when evaluation details change. Read OpenAI’s October GPT-6 system card.

The model name does not identify the surface

OpenAI states that October GPT-6 Sol and Luna went to ChatGPT, while Work and Codex still used their September versions when this card was published. A ChatGPT conversation, Work session, and Codex task are not interchangeable tests just because each carries a GPT-6 family label. Record the surface and release snapshot. Verify the served model when the product exposes it.

Recent complaints describe friction, not a controlled result

Recent English- and Spanish-language Reddit posts describe memory or context problems, repeated instructions, unwanted tone or familiarity, and extra correction work. One Spanish-language Plus user says a Work task did not produce the expected result and that the user then tried Claude. Another thread includes both complaints about joking or awkward style and a report that GPT-6 infers imprecise intent better than GPT-5.6. These posts name real user costs, but they include no matched output archive, verified route, or denominator. They do not establish cause or prevalence.

Today’s service check is a separate lane

OpenAI Status is investigating elevated errors in hosted Code Reviews and Security Reviews on private GitHub repositories. The incident page did not yet mark a component as affected. It is not a report of GPT-6 Chat answer quality. Anthropic Status listed its displayed components as operational and no incidents on 26 or 27 October when checked. Provider status helps classify availability; it does not grade completed answers.

A public tracker cannot settle this comparison

AgentCheck’s public index runs 20 fixed, deterministic prompts daily. Its 26 October page shows stable exact-golden scores for selected Anthropic models, but says it has not measured Google or OpenAI. It checks simple answer consistency, not GPT-6 or a long ChatGPT conversation. Stable results on that panel cannot confirm or rule out the GPT-6 safety and context reports.

What would change the call

An independent comparison should pin the ChatGPT release and surface, use a repeatable task set, record conversation state and settings, and separate failed requests from completed answers. Score factual correctness, task completion, safety, and the time or corrections a user needs. Repeat the tasks to measure normal variation. A provider test can identify a narrow behavior change; a user can decide whether it matters only when the same work is tested on the system they use.

See the current verdict · Browse community reports · Compare evidence lanes · Read the method