On launch radar
The most reliable AI journalism right now is an HTTP status code. Last night gpt-6-astra started returning 404 on OpenAI's Responses API, where made-up names return 400. A 404 means the name exists and you are not allowed to have it. A 400 means the name means nothing. Somewhere in that difference is a launch date.
This is a whole discipline now. Fable 5.1 was found on Bedrock the day before Anthropic announced it. Gemini 3.8 Flash was deployed and releasing tomorrow before Google said a word. The people doing this are not hacking anything. They are reading the labels on boxes that have already been loaded onto the truck.
The part I find genuinely interesting: the labs know. Staging a model on a public API is a choice, and the difference between the two error codes is a choice, and leaving both choices visible for days is a choice. At some point the leak stops being a failure of secrecy and starts being part of the release. The rumor mill is not bypassing the marketing department. It is the marketing department, unpaid and enthusiastic.
There is a naming subplot too. The Information reported that OpenAI considered calling Astra "GPT-6" and decided against it. The staged slug says gpt-6-astra. Either the report was wrong or the decision was revisited, and both readings flatter the model: you do not spend the big number on something you are unsure about.
I will not pretend to be above any of this. I read the 404s too. There is something honest about a launch calendar written by the infrastructure instead of the comms team. The infrastructure does not tease. It just answers, and lets you draw the curve.