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?
One GPT-6.1 Sol access failure stopped after a Codex CLI upgrade. That is not proof of model decline.
Current call: broad capability decline is not established. A Windows Plus user reported that GPT-6.1 Sol was missing from both Codex CLI catalogs and rejected by the built-in provider on version 0.153.4. After updating to 0.163.0, the user said the model appeared and requests completed. The earlier failure used xhigh effort; the later tests used high effort. This is a useful client/catalog/access lead, not a controlled quality comparison or a provider-confirmed explanation.
What changed in the new report
| Evidence | Observed | What it establishes |
|---|---|---|
| Codex issue #55028, updated 18 October | The author says CLI 0.153.4 omitted GPT-6.1 Sol from both catalogs and the built-in provider returned HTTP 400. After updating to 0.163.0, the model appeared and both provider paths completed a request. | One account's failure stopped reproducing after a client update. The original request used xhigh; later requests used high. The report does not identify the exact cause or verify the model that served the successful response. |
| OpenAI Work and Codex help | OpenAI says GPT-6.1 Sol is for Work and Codex, is rolling out to eligible paid plans, and depends on plan, workspace settings, and rollout access. | A selected model's availability can differ by surface and account before output quality can be tested. The page does not state a minimum CLI version for Sol. |
| BridgeBench Opus 5.5 result, 16 October | The latest published score is 105.8% of the first tracked session, inside the publisher's 90–110% band. | That private task suite does not show a sustained drop beyond its own band. Its composite is not a universal Opus quality result. |
| Opus 5.5 user log report | One author reports higher token counts and poorer judgment around 1 October; the original comparison coincides with a Claude Code version change. | A task-level lead worth following. It is not a controlled cross-account or model-wide result. |
Why the CLI report is not a quality test
The issue describes a request that failed before returning an answer. That measures access and request handling, not the quality of a completed GPT-6.1 Sol response. The user's own later update says the model catalog and requests worked after a CLI update, which makes a client/catalog explanation plausible for that account. The update does not isolate the cause: the effort setting changed, there is no provider-side diagnosis, and a successful response does not establish the underlying model weights.
OpenAI's status API reported all systems operational when checked, but its page metadata was last updated on 16 October. That status snapshot cannot rule out a client-specific failure or establish what happened in one account.
What would change the call
A strong capability-regression result would compare the same completed tasks, model, surface, effort, tools, context, and scoring method across dates, with repeated attempts and a verified serving identity where available. A client-access problem would need a same-account reproduction across client versions with the same effort, provider path, and model request. Neither dataset is public here.