OpenAI released GPT-6 Astra last week, and according to early assessments, it represents a substantial improvement over GPT-5.6 Sol across multiple domains. The model demonstrates particular strength in mathematics, coding, and 3D rendering tasks.
On independent benchmarks, Astra achieves noteworthy scores. According to the Artificial Analysis Intelligence Index, it ranks at the frontier of current models, though it does not dramatically outpace competitors on general agentic coding tasks. One significant achievement cited is a score of 99.9% on the ARC-AGI-3 benchmark, which measures logic puzzle solving and generalization ability—compared to GPT-5.6 Sol’s 7.8%.
A notable capability of Astra is its computer-use functionality, particularly with graphical user interfaces. The model can operate software on local computers through the ChatGPT app, performing tasks like drawing images in MS Paint using mouse controls. While not the first model to demonstrate computer-use capability, Astra’s implementation appears more mature than earlier efforts.
According to recent reporting, OpenAI purchased tens of thousands of Mac Minis and Mac Studios to train computer-use capabilities. The training workflow involves providing the model with screenshots of the macOS interface, having it predict mouse and keyboard actions, executing those actions on the hardware, and capturing new screenshots to create feedback loops. This approach allows Astra to learn to interact with graphical interfaces during training.
The model’s improvements in rendering and animation tasks have made for particularly compelling demonstrations on social media, ranging from 3D modeling in Blender to virtual open house tours. Astra appears especially capable at tasks requiring graphical interaction and visual output, leapfrogging its predecessor in categories including writing, math, coding, and more.
The use of Mac hardware for training reflects a broader trend toward making language models capable of everyday computer tasks. As developers refine computer-use capabilities across both the model and agent framework layers, expectations are that LLMs will become increasingly accessible for non-technical users performing standard computer workflows.
Key facts
- GPT-6 Astra achieves 99.9% on the ARC-AGI-3 benchmark, up from GPT-5.6 Sol’s 7.8%
- The model demonstrates exceptional performance on 3D rendering and animation tasks relative to other models
- Astra can operate graphical user interfaces on local computers, including tasks like drawing in MS Paint
- OpenAI purchased tens of thousands of Mac Minis and Mac Studios to train computer-use capabilities
- Computer-use training involves iterative cycles of task prompts, screenshots, predicted actions, and execution feedback