GPT-6 Astra can turn photos into 3D models and playable environments

OpenAI released GPT-6 Astra this week, rolling it out first to a limited set of organizations before extending access to ChatGPT Plus, Pro, Business, and Enterprise users, and to developers through the OpenAI API, Microsoft Azure, and AWS Bedrock. Among the capabilities OpenAI highlighted at launch, 3D modeling and spatial reasoning stood out as one of the more direct points of interest for design and manufacturing workflows.

To measure the model’s grasp of 3D geometry, OpenAI introduced BenchCAD, a benchmark that tests whether a model can reconstruct a 3D object from a set of multi-view renders by writing CAD code rather than by generating a mesh directly. Working with tools, GPT-6 Astra scored 95.9% on the benchmark’s mean voxel IoU metric, a measure of how closely the reconstructed geometry overlaps with the original object. That compares with 83.3% for GPT-5.6 Sol, OpenAI’s prior model, and 84.3% reported for Anthropic’s Claude Fable 5.1. OpenAI also reported that Astra reached this score at roughly 43% lower estimated API cost than Sol and 86% lower than Fable 5.1 in the configurations tested.

OpenAI paired the benchmark result with a demonstration of the model modeling a house in Blender and converting it into a walkable scene in Unreal Engine 5, describing the workflow as a way for designers and clients to explore a layout and move through a space before it gets built. The same computer-use capabilities extend to game development: in demos credited to designer Pietro Schirano, Astra built a kart racing game and a set of spaceship stills, generating graphics, gameplay, and motion for people without programming experience.

For readers tracking additive manufacturing and product design, the BenchCAD result is the one worth watching. A model that reconstructs editable CAD geometry from ordinary photos, rather than a static mesh, points toward faster reverse engineering of physical parts, quicker iteration between a scanned reference and a printable file, and design tools that generate working geometry instead of just images.

Whether Astra’s output holds up against the tolerances additive manufacturing workflows require is a separate question from the benchmark score, and one that will depend on how the model performs once designers start feeding it real parts rather than test renders.