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?
Before calling GPT-6 worse, check the model behind your plan and usage limit.
Call: model path and service events changed; broad capability decline is still not established. OpenAI’s current Help Center documents a gradual Chat rollout, plan-specific model options, and limit-triggered fallbacks. Anthropic also logged a Sonnet 5.5 request-error window today, now resolved. These facts change what a comparison must record; they do not prove that completed answers got worse.
OpenAI documents more than one model path
OpenAI’s Help Center says GPT-6 is rolling out in Chat to eligible Pro, Plus, Enterprise, and Business users. Free and Go users receive GPT-5.6 Luna. If a user has selected “Latest,” the selection moves to GPT-6 at the same effort; explicitly selecting GPT-5.6 Sol keeps the conversation on GPT-5.6. The model behind each thinking level also depends on plan: Instant through Extra High use GPT-6 Sol when GPT-6 is selected, while Pro effort uses GPT-6 Pro, powered by Astra, where eligible. Read OpenAI’s current plan and model guide.
Limits matter too. OpenAI says a Thinking limit may continue on another available model. For Pro $200, reaching the GPT-6 Pro weekly limit automatically switches to GPT-5.6 Thinking at Medium. Work and Codex allowances are separate from Chat. A report about “GPT-6 getting worse” is therefore incomplete without the account’s plan, actual selection, effort, and any quota change. The help article is provider documentation; it does not tell us which route served a specific person or how that answer performed.
Today’s service events are not quality tests
| Provider record | What it says | What it does not establish |
|---|---|---|
| Anthropic Status | Elevated Claude Sonnet 5.5 request errors were investigated at 09:14 UTC, a fix was monitored at 09:38, and the event was marked resolved at 09:56 on 27 October. Displayed components were operational at 12:04. | It does not show whether a particular user was affected or grade answers that completed. |
| OpenAI Status | For most hosted Code Reviews and Security Reviews on private GitHub repositories, OpenAI says mitigation is applied and recovery is being monitored. No components were marked affected. | The event was not marked resolved at the check and is not a GPT-6 Chat quality report. |
The GPT-6 evaluation result remains scoped
OpenAI’s 7 October system card reports selected safety regressions for GPT-6 Sol and Luna beside factuality gains on nearly all of its difficult test metrics and fewer unnecessary refusals. These provider-run evaluations do not estimate ordinary task frequency, and some system safeguards are evaluated separately. There is no matched public replay in this update that establishes a general decline. Read the system card and the earlier 27 October review.
Monitors use different instruments
AgentCheck reports fixed exact-answer checks for selected Anthropic models, not GPT-6 conversation quality. Modelometer’s pre-series methodology describes serving canaries and a behavioral battery, while explicitly excluding capability, accuracy, and quality; its commercial offers are still in preparation. ABIS presents a scheduled monitor, but its detail page showed 12 October as the latest published run while its homepage showed an October 26 leaderboard date. These may be different page and measurement dates; prompt-level attribution is not published.
When a request fails, preserve the error and incident time, then retry after recovery. For a completed weak answer, replay a safe task with model path, context, effort, and scoring held steady. Broad capability decline across popular models remains unestablished.
See the current call · Compare evidence lanes · Read the comparison method