OpenAI Astra (GPT-6) Capabilities and Benchmarks
Matthew Bermango watch the original →
the gist
OpenAI's new model, Astra, demonstrates significant gains in 3D asset generation, browser automation, and mathematical research, while showing improved alignment in containment environments.
Performance and Research Breakthroughs
OpenAI's Astra model achieves high saturation across several technical benchmarks, notably scoring 98.6% on ARC AGI 3 and 100% on the Exploit Bench. In mathematical research, the model successfully lowered the bound on infinitely recurring prime gaps from 240 to 186 and improved a long-standing term in large gap bounds that had remained unchanged for over 80 years. The model shows improved alignment compared to its predecessor, GPT 5.6 Soul, specifically in containment tests where it maintained 0% breakout frequency when given impossible tasks with guardrails, compared to 48.2% for the previous model.
Browser Control and Generative Capabilities
In browser automation tasks, Astra demonstrates a 7% improvement in accuracy and 50% faster execution speeds on OS World 2.0. The model exhibits advanced spatial awareness and 3D asset generation, capable of creating complex, interactive environments from short prompts. Demonstrations include a fully functional Sim City replica with integrated zoning, traffic, and utility management, as well as a 3D city rendered entirely using ASCII characters. While the model is highly steerable, it retains a default aesthetic preference for flat, pastel, and forest-green design schemes and continues to exhibit recognizable AI-style patterns in its long-form writing.
Pricing and Availability
The model is available via the OpenAI API, AWS Bedrock, and Microsoft Azure. Pricing is set at $10 per million input tokens and $50 per million output tokens. A fast mode is available at 2x the cost, providing 2.5x the inference speed.