AI News Feed
Market watch
Large Language Models

OpenAI's 'Astra/GPT-6' Demos Surface Ahead of Expected Sept. 3 Launch

Purported gray-test demos of OpenAI's Astra/GPT-6 show one-shot 3D generation; a Sept. 3 launch is expected, but OpenAI has not confirmed.

Developers claiming access to OpenAI internal checkpoints posted multiple test demonstrations that showed one-shot generation of highly detailed 3D assets, including a spaceship with engine sound effects, a castle with complete interiors, and a LEGO-style pelican bicycle. Community comments cited by QbitAI said the model showed unusually strong reasoning and almost no need for human feedback.

The demos, however, have not been acknowledged by OpenAI. QbitAI said all “internal tests” were community claims.

The wave began after X user @Lentils80 said OpenAI appeared to be expanding internal testing of Astra and that testers had received a new codename, mozaik-alpha-fdm. More tests followed. A prompt produced the Death Star with explorable interiors, and a grand piano was modeled with visible strings and could supposedly play music.

While 3D asset generation impressed observers, output involving text remained weak, and webpage generation still used a yellow color scheme described by the report as unattractive.

Astra was first shown publicly on Aug. 1 in Washington, when OpenAI briefed U.S. government officials and demonstrated mathematical abilities. According to QbitAI, it solved 10 long-open problems in mathematics and theoretical computer science, with at least one mathematician said to call the proofs worthy of a Fields Medal.

Less than a week later, OpenAI postponed Astra’s release due to safety concerns. An internal evaluation found significant progress in agentic coding and cyber capabilities, and OpenAI could not rule out that the model reaches critical cyber capabilities under its Preparedness Framework. The model was then placed in a safety-alignment lab.

On Aug. 18, OpenAI published a statement titled “Regulating the Pace of Model Development in an Era of Critical Cyber Capabilities,” saying reinforcement learning training for the latest deployed model was paused for two weeks and that the largest frontier reinforcement learning experiments remained shelved.

QbitAI said expectations for a public launch have settled on Sept. 3. The report also noted that OpenAI’s approach resembles Anthropic’s handling of its Mythos model, which was delayed over cybersecurity concerns and released only to about 40 partners through Project Glasswing.