Multi-Vendor LLM Catalog Probe: Catalog-Inference Layer Asymmetry and DashScope Aggregation Evidence (2026-06-20)

Single-timestamp /v1/models endpoint snapshots across 6 major LLM vendors (Anthropic, OpenAI, Google Gemini, xAI, DeepSeek, Qwen/DashScope; 333 models total) captured 2026-06-20 12:50 JST. Documents (1) DashScope hosting DeepSeek v4-flash, v4-pro and GLM 5.1 alongside the Qwen family, evidencing Alibaba's aggregation of third-party Chinese LLMs onto a single compatibility layer, and (2) the catalog-vs-inference asymmetry of Anthropic's post-takedown response to Claude Fable 5 / Mythos 5 (Mythos 5 fully removed from all catalogs; Fable 5 retained in Anthropic's catalog but returning HTTP 404 at the inference layer). Companion temperature-parameter deprecation probe confirms Opus 4.8 rejects temperature values in [0,1) with HTTP 400.