OpenAI has expanded internal testing of an unreleased model called Astra, running a checkpoint referred to online as “mozaik-alpha-fdm,” according to TestingCatalog. The outlet’s account rests on clips posted to X by accounts with no stated OpenAI affiliation, credited to a leaker going by Lentils and to a Discord server called zAI. TestingCatalog did not obtain the material directly from OpenAI, and the company has not confirmed the checkpoint name.
The circulating clips show a voxel castle, a GTA 2 style game, and several detailed websites, each said to come from one prompt under a slower reasoning setting the leaks label “Max effort.” That is a small set of selected demos, not a benchmark, and cherry picking is standard practice for this kind of preview.
OpenAI has separately acknowledged an upcoming model named Astra and said internal evaluations flagged major gains in coding automation and security work, serious enough that some workloads were paused pending stricter safeguards, per TestingCatalog. That statement confirms Astra exists as a project. It says nothing about whether the leaked clips reflect what eventually ships. GPT-6 remains an unconfirmed public name for the model, TestingCatalog noted.
Reporting by TestingCatalog, published August 29, 2026.
