OpenAI’s GPT-6 Astra scored 95.9% on BenchCAD’s mean voxel IoU metric, a benchmark that tests whether a model can turn multi-view photos of an object into editable CAD code rather than a static mesh. That’s up from 83.3% for OpenAI’s prior model, GPT-5.6 Sol, and ahead of the 84.3% reported for Anthropic’s Claude Fable 5.1. OpenAI also said Astra hit that score at roughly 43% lower estimated API cost than Sol and 86% lower than Fable 5.1 in the tested configurations.
The distinction the benchmark measures matters for manufacturing specifically: a mesh is a shape you can view, while CAD code is a shape you can edit, parametrize, and open in standard design software, which is the format printable files start from. OpenAI is rolling Astra out first to a limited set of organizations, with ChatGPT Plus, Pro, Business, and Enterprise access, plus API access through OpenAI, Microsoft Azure, and AWS Bedrock following in the coming days.
OpenAI paired the benchmark result with a demo outside BenchCAD itself: Astra modeling a house in Blender and converting it into a walkable Unreal Engine 5 scene. BenchCAD tests geometric overlap against clean, multi-view renders built for evaluation, a step removed from scanned or photographed real-world parts and the dimensional tolerances that decide whether a file actually prints.
The result lands in a space specialized tooling already occupies. Backflip, founded by Markforged’s Greg Mark and David Benhaim, says its newest model cut part digitization costs from roughly $1,500 to $10 per part by building CAD “the way engineers actually do.” Creality has partnered with KVS Ltd on an integrated scan-to-CAD workflow for reconstructing damaged parts for CNC machining or 3D printing.