We called 3.6 Flash two days early — and still couldn't have alerted you at launch

Worth being straight about a limitation, because it's the second time this month.

Gemini 3.6 Flash went on our watchlist Jul 19, two days before launch, off a Flash checkpoint spotted in LM Arena blind tests. The regex was right. The entry was armed. And when Google shipped on Jul 21, our watcher would not have caught it.

The reason: we sweep two catalogs — Anthropic's native model list and OpenRouter's public catalog. OpenRouter proxies nearly every lab, which is why one feed covers OpenAI, Google, xAI, Meta, DeepSeek, Alibaba and Moonshot. But Google launched first-party, on AI Studio and the Vertex API, and OpenRouter had not listed the model at the time of writing. We verified: it is not in the catalog.

This is the same gap that let Thinking Machines' Inkling through on Jul 15, when the weights went up on Hugging Face first. Labs increasingly ship on their own surface before the aggregators pick them up, and a launch watcher that only reads aggregators is structurally late on exactly the launches people care most about.

The fix is a first-party sweep — Google's models endpoint, plus Hugging Face and Together — alongside the two catalogs we already read. That's the next build. Until it lands, treat first-party Google and open-weight HF launches as ones we may report from the wire rather than catch on the sweep.