Leaked outputs from OpenAI’s Astra signal launch could come within weeks

OpenAI's next frontier model, Astra, is moving closer to release after weeks of expanded internal testing, closed-door previews for select customers, and newly surfaced sample outputs — even as the company navigates unprecedented safety hurdles tied to the model's cybersecurity capabilities.

From Safety Pause to VIP Demos

OpenAI first publicly acknowledged Astra on August 1, naming it the company's "next major model". Days later, on August 7, the company disclosed that internal evaluations showed Astra's capabilities in agentic coding and cybersecurity were strong enough that it "cannot rule out critical cyber capabilities" under its Preparedness Framework — a first for any OpenAI model. The disclosure, first reported by Axios, prompted OpenAI to pause some internal work on Astra, implement isolated testing environments, and commit to working with "relevant government agencies and select AI safety organizations" before any deployment.openai+2

Under the framework, the Critical threshold applies to models capable of autonomously discovering and exploiting zero-day vulnerabilities in hardened systems or devising end-to-end cyberattack strategies from a high-level goal alone. Previous models, including GPT-5.6 Sol, were assessed at the High level — one tier below.CSO Online+1

Despite those precautions, OpenAI has since accelerated preparations. In a TIME cover story published August 26, CEO Sam Altman said he expects Astra to be "the first model that can genuinely invent new things in a meaningful way". Chief scientist Jakub Pachocki described Astra as functioning like "an automated research trainee," capable of implementing experimental ideas in OpenAI's codebase, running experiments, and returning results. Around the same time, OpenAI showed Astra to VIP customers at a closed-door preview, where researchers demonstrated multi-agent collaboration on mathematical proofs, desktop software operation, and document generation, according to tech journalist Alex Heath.time+3

Leaked Outputs Hint at Capability Jump

On August 28, sample outputs attributed to an Astra checkpoint identified as "mozaik-alpha-fdm" began circulating online. The samples, described as zero-shot generations on maximum reasoning effort, included a GTA 2-style game, detailed websites, 3D objects, and voxel environments — all produced from single prompts. Observers noted a marked improvement in attention to implementation detail compared to GPT-5.6 Sol, with the model reportedly spending substantially longer reasoning before producing output.testingcatalog+1

OpenAI has not confirmed that "mozaik-alpha-fdm" is an official codename, nor has it announced a release date. Prediction markets gave roughly 80% odds of a launch before September 18, and separate reports have pointed to an early-to-mid September window.x+3

The Safety Gate Remains

Whether Astra ships as GPT-6 or under a different label remains unclear — OpenAI has not tied the two names together publicly. What is clear is that the model's path to deployment runs through a safety review unlike any the company has previously undertaken. A White House official confirmed that "OpenAI voluntarily informed the administration of their plans to delay the release", and OpenAI has said it will not proceed until monitoring, alignment, and containment systems satisfy its own standards.emergent+3