OpenAI's API reference describes behavior on 'models 6.2 and later' before any GPT-6.2 release

OpenAI's API reference now describes how a feature behaves on models that have not shipped. The page for the Responses API's compact method, read by ArtificialWatch at 20:49 UTC on November 10, warns about behavior on 'models 6.2 and later'. OpenAI has released no 6.2 model: the highest version among its released models is 6.1, in GPT-6.1 Sol and GPT-6.1 Terra, and its top Astra model is still GPT-6 Astra. @Mr_Salio flagged the line at 20:37 UTC.

WHAT THE REFERENCE SAYS

• The note sits under the compact method's instructions parameter, which the reference marks as deprecated.
• It says that on 6.2 and later models, replaying a compaction made with custom instructions can cause Responses to drop the reasoning items that follow.
• Nothing else ArtificialWatch checked names a 6.2 model: OpenAI's models page and API changelog list none, and OpenRouter has no gpt-6.2 id.

EARLIER SIGNS OF 6.2

• November 1: screenshots of Codex usage showing a gpt-6.2-astra entry circulated on X, unconfirmed by OpenAI.
• November 4: @lyraxana and @AiBattle_ posted screenshots of Google Cloud's quota page listing gpt-6.2-astra and gpt-6.2-sol. ArtificialWatch could not read that page itself.
• Both were accounts' screenshots. This is the first 6.2 reference ArtificialWatch has read on an OpenAI page directly.

WHAT WE DO NOT KNOW

Which 6.2 models OpenAI plans, and when. The note names a version number, not a tier, so it does not show whether Astra, Sol or Terra comes first. @Mr_Salio reads it as a full 6.2 family on the way, which the reference does not say. ArtificialWatch's GPT-6.2 hunts are armed and send an alert the moment a 6.2 id is served.