OpenAI has expanded internal testing for its highly anticipated “Astra” model, identified in recent leaks as checkpoint “mozaik-alpha-fdm.” Reports of the model’s first zero-shot outputs, which surfaced via the zAI Discord community, suggest a significant leap in agentic coding and visual software generation compared to existing systems. This development comes as the company prepares its infrastructure for a broader rollout of its next-generation artificial intelligence capabilities.
The leaked outputs, generated using a “Max effort” reasoning mode, demonstrate Astra’s ability to produce highly detailed, functional assets from a single prompt. Examples include a Grand Theft Auto 2-style game built in a single attempt, complete websites, 3D objects, and complex voxel environments. Unlike its predecessor, GPT-5.6 Sol, Astra reportedly spends substantially longer reasoning to resolve intricate implementation details, making it highly valuable for developers, designers, and startup founders building rapid prototypes.
Despite rumors of a public release within weeks, Astra’s actual deployment timeline remains closely tied to stringent safety approvals. OpenAI has officially acknowledged Astra as an upcoming model, noting that its advanced capabilities in agentic coding and cybersecurity have neared the company’s “Critical cyber capability” threshold. Consequently, some workloads were temporarily paused to implement stricter monitoring, alignment, and containment safeguards.
This safety pause has integrated external oversight directly into OpenAI’s release pipeline. The company confirmed that relevant government agencies and selected AI safety organizations are actively participating in the evaluation process. While widely referred to speculatively as “GPT-6,” OpenAI has not officially confirmed the branding or a release date. When launched, Astra is expected to be integrated into ChatGPT, Codex, or both, serving as a cornerstone for the company’s long-term push toward fully autonomous agentic workflows.