Shaduf.
AI model watch · 22 September 2026 · 12:00 UTC

An outage and a surface split. Still no broad decline.

Claude had a resolved multi-model incident today. OpenAI says its recent Sol update was Chat-only. Both facts sharpen the test without proving that popular model weights got worse.

Call: no broad cross-model core-capability decline is proven. Treat today’s Claude event as availability evidence. Treat Chat-versus-Work complaints as a surface replay lead. Do not turn either into a weight verdict.

What changed today

Claude · availability
00:50–02:10

Anthropic reported elevated errors for Fable, Mythos, and Opus across Claude surfaces, then resolved the incident at 02:35 UTC. Retry or wait is the proportionate move.

OpenAI · product boundary
Chat ≠ Work

OpenAI says the 6 August Sol update changed Chat, while the version powering Work and Codex was not changing in that release. The surface belongs in the test identity.

Community · system lead
Replay it

Reports describe fast or shallow Chat responses, file/tool failures, and Chat-versus-Work differences. They are detailed intake, not a population estimate.

Measurement · guardrail
Judge too

Judge version and severity can move scores. The latest visible public drift board is also dated 13 September, so alert age matters.

Reader move

Before switching, record the fixed task, surface, requested and served model, route, plan, context, configuration, judge or harness, availability state, task budget, warning or stop state, completed artifact, correction burden, allowance movement, and control result. Then choose retry, wait, pin, stop, switch, or replay.

Read the 22 September report · See the receipt method

Search published pools, pages, reports, and evidence.