Latent Space has released GPT-6 Astra, a Stargate model designed to perform a wide range of AI engineering tasks. Initial testing indicates Astra’s performance surpasses Fable 5.1 across multiple benchmarks, including achieving 97.6% accuracy on FrontierMath and 99.9% on ARC-AGI-3. The model’s capabilities extend to managing complex agent systems, handling data labeling, and optimizing pipeline performance.
Astra’s operational cost is estimated at $6 per hour, based on a token processing rate of 33 tokens per second and a maximum rate of $50 per million tokens. This efficiency is confirmed by independent analysis, placing Astra competitively against models like Sol and Fable. The system is designed to manage up to 20-50 parallel agents, controlled by a central Astra agent, allowing for efficient task execution.
Key functionalities include automated monitoring of runs, initiation and termination of execution waves, and the creation of benchmarks. This functionality mirrors the work typically performed by a junior AI engineer, reducing operational costs significantly. Furthermore, Astra can generate personalized Arena.ai clones for prompt tuning, model selection, and preference model alignment.
Latent Space’s experimentation with Astra has resulted in a diverse range of projects, including training game AI, automating personal finance tasks, and redesigning internal tools. The team’s ambition, fueled by Astra’s capabilities, has led to rapid development and deployment of multiple SaaS tools and internal systems. Source: https://www.latent.space/p/astra
