Anthropic, OpenAI, Meta and xAI models were live before the announcement in 9 of 11 launches; nine other labs announced first every time

In 9 of the 11 launches from Anthropic, OpenAI, Meta and xAI that ArtificialWatch timed to the minute since July, the model was answering on a public API before the lab posted about it. In all 9 timed launches from other labs, the announcement came first, by anywhere from an hour to 68 days.

LIVE BEFORE THE POST

- Anthropic, 5 of 5: every Claude launch we timed, from Claude Opus 5 in July to Claude Haiku 5.5 in October, was live 6 to 11 minutes before @claudeai's post.

- OpenAI, 2 of 3: GPT-6.1 Terra 14 minutes and GPT Image 2.6 12 minutes before the post. GPT-6 Astra went live about 24 hours after it.

- Meta, 1 of 1: Muse Spark 1.3 answered about 2 hours 47 minutes before @AIatMeta's post.

- xAI, 1 of 2: Grok 4.7 was live 28 minutes before the post; Grok 4.6 came 22 minutes after.

ANNOUNCED FIRST

Moonshot's Kimi K3.1 Pro went live an hour after its post, Mistral Large 4 about 1 hour 36 minutes after, Ant Group's Ling 3.1 Flash about 1 hour 48 minutes after, and Inception's Mercury 2.5 about 14 hours after. The longer waits: Reflection AI's Beam 11 days, Google's Gemini 4 13 days, Alibaba's Qwen 3.8 Max 15 days, MiniMax M3.1 Flash 22 days and Tencent's Hy4 68 days.

WHAT IT MEANS FOR ALERTS

When a model is live before the post, a sweep of the labs' APIs can report it before the lab does; the median lead in this record is 11 minutes. When the post comes first, the useful alert is the day the model actually answers, which for most of the labs above came days later.

HOW WE COUNTED

From ArtificialWatch's lead-time record (artificialwatch.com/lead-time, also at /data/lead-time.json): the minute our sweep first saw each model answer on a public API, against the time of the lab's own announcement post, with each post linked. 21 launches from July 24 to November 4; GPT-6.1 Sol is left out because OpenAI introduced it in a livestream that began before its post. Most labs have a single timed launch here, so their pattern rests on one data point.