Anthropic's last five public models were live on its API 6 to 11 minutes before it announced them

ArtificialWatch's sweep found each of Anthropic's last five public models live on Anthropic's API before Anthropic announced it on X, by 6 to 11 minutes each time. No other lab in ArtificialWatch's launch record did it more than once.

THE FIVE (UTC, seen live on the API, then the @claudeai post)

- Claude Opus 5, July 24: seen 16:51:33, announced 16:59:51, 8 min 18 s early. - Claude Fable 5.1, September 1: seen 17:52:22, announced 18:03:14, 10 min 52 s early. - Claude Opus 5.5, September 22: seen 16:24:44, announced 16:31:01, 6 min 17 s early. - Claude Sonnet 5.5, September 28: seen 17:54:45, announced 18:03:31, 8 min 46 s early. - Claude Haiku 5.5, October 7: seen 17:51:44, announced 18:01:16, 9 min 32 s early.

The median gap is 8 min 46 s. Mythos 5.1, which Anthropic shipped on September 1 to verified organizations only, is not on the public API and is not in the record.

THE OTHER LABS

The record holds 15 launches from July 24 to October 26. Seven went live before they were announced: Anthropic's five, Meta's Muse Spark 1.3, which ArtificialWatch saw 2 h 47 min before @AIatMeta's post on September 2, and OpenAI's GPT-6.1 Terra, seen 13 min 43 s before @OpenAI's post on October 26. Meta's blog post carries a date but no time, and OpenAI's has no machine-readable time, so neither can show an earlier announcement either way.

The other eight were announced first:

- OpenAI announced GPT-6 Astra on September 3 with API access rolling out over the following days. ArtificialWatch saw it live 23 h 38 min after the announcement, and 62 minutes before OpenAI's post saying it was also live in the API. - Google announced Gemini 4 on September 30 for trusted testers only. ArtificialWatch saw it 26 minutes before Google's October 13 general-availability post. - MiniMax announced M3.1 Flash on September 27 as a subscription-only preview. ArtificialWatch saw the full release 5 minutes before MiniMax's October 19 launch post, though OpenRouter had listed it 8.7 hours before ArtificialWatch's sweep caught it. - Mistral Large 4 came 1 h 36 min after Mistral's blog post and 30 minutes after its post on X. - Ant Group's Ling 3.1 Flash came 1 h 48 min after its announcement, and Inception's Mercury 2.5 13 h 56 min after. - Alibaba announced Qwen 3.8 Max as a preview on July 19; ArtificialWatch saw the full release 2 h 6 min after Qwen's August 3 post. - Reflection AI introduced Beam on October 5 with no availability; ArtificialWatch saw it 46 minutes after the October 16 availability post.

OpenAI's GPT-6.1 Sol is left out: OpenAI introduced it in a DevDay livestream that began before its post, so the post may not be the first announcement.

HOW IT IS MEASURED

First seen is the first time ArtificialWatch's sweep, which checks the providers' own APIs and public model catalogs every 60 seconds, found the model live. The announcement time is the lab's post on X, read to the second, or the blog's machine-readable publish time where the blog came first. Anthropic's model pages show a date but no time, so an earlier untimed announcement there cannot be ruled out. Every row was checked by hand. Launches after October 26 have not been added yet. The full record, with a link to every announcement, is at artificialwatch.com/lead-time, and as open data at artificialwatch.com/data/lead-time.json and .csv.