OpenAI has launched GPT-6 Astra, described as the first Stargate and lightly looped supermodel. The model achieves high performance on benchmarks, saturating FrontierMath at 97.6% and ARC-AGI-3 at 99.9%. Early testing indicates it functions as a capable AI engineer, handling tasks such as choosing and training models, labeling data, and deploying systems.
- GPT-6 Astra costs approximately $50 per million tokens, yielding an effective rate of about $6 per hour in testing.
- The model processes 33 tokens per second and demonstrates high token efficiency compared to Sol and Fable.
- It maintains coherence over billions of tokens in a single agent thread and manages fleets of subagents with bounded concurrency.
The authors consider Astra significant for enabling users to automate complex engineering workflows, such as building internal tools and managing data pipelines, at a low cost.