Shaduf.
AI model watch · 24 September 2026 · 12:00 UTC

The rollout moved before the verdict did.

OpenAI put GPT-6 Sol and Luna in Work and Codex, not ordinary Chat. Work then had a declared incident, and a community trace shows client and server model labels disagreeing. The evidence points to surface and provenance, not a broad collapse.

Call: no broad cross-model core-capability decline is proven. The strongest current explanation for the visible change is a moving product system: release boundary, execution surface, route, attribution, capacity, and allowance. Replay before you switch.

What changed today

OpenAI rollout
Work ≠ Chat

GPT-6 Sol and Luna launched in Work, Codex, and the API but not ordinary Chat. OpenAI says production Chat can differ from API or research evaluation because tools and system prompts differ.

Availability
Resolved

OpenAI recorded elevated ChatGPT Work errors across paid and education plans. Claude also recorded a resolved multi-model incident around the same rollout window.

Identity
Fields disagree

A self-reported ChatGPT Web trace shows a client model slug changing after completion while server and native attribution remain GPT-5.6. Treat it as a provenance lead, not proof of fallback.

Counterweight
Opus 5.5

Anthropic reports a positive release, while early users disagree by task. That is a counterweight to any universal decline story, not a shared benchmark.

Reader move

Before switching, capture release boundary, surface, execution class, requested model, provider/server/native/client/final model identity, route, system prompt and tools, context and hidden I/O, judge or harness, availability, task budget, warning or stop state, artifact, correction burden, and control. Then choose retry, wait, audit, stop, switch, or replay.

Read the 24 September report · See the receipt method

Search published pools, pages, reports, and evidence.