OpenAI released its first Stargate and lightly recurrent GPT-6 Astra model on September 3, 2026. This model performed exceptionally well in various benchmark tests, achieving scores of 97.6% for FrontierMath and 99.9% for ARC-AGI-3. After being trained and tested with over 20 billion tokens, Astra was proven to possess the capabilities of a fully automated AI engineer, including selecting training models, annotating data, maintaining pipelines, debugging systems, and directing evaluation sub-agents. In terms of operating costs, when running at a rate of up to 50 dollars per million tokens, its speed was approximately 33 tokens per second; although managing 20 to 50 sub-agents in parallel increased overhead, tests showed that its hourly operating cost could be kept below 6 dollars. Currently, this model outperforms Sol and Fable in token efficiency and has been verified as one of the fastest intelligent models besides Spark 1.3.
OpenAI today released its first Stargate and lightly recurrent GPT-6 Astra models. These models outperform Fable 5.1 in multiple metrics, including fully covering FrontierMath (97.6%) and ARC-AGI-3 (99.9%). After testing with over 20 billion tokens consumed, the model was proven to be a new type of fully automated AI engineer, capable of selecting training models, annotating data, maintaining pipelines, debugging systems, and commanding evaluation sub-agents. In terms of cost, operating at a rate of up to $50 per million tokens, Astra can achieve a speed of approximately 33 tokens per second. Although managing 20 to 50 sub-agents in parallel significantly increases overhead, testing shows that its hourly cost can be kept below $6. Currently, this model outperforms Sol and Fable in token efficiency and has been verified as one of the fastest intelligent models besides Spark 1.3.