GPT-6 Astra Weight Adjustment Causes Explosive Growth? Controlled Experiments Reveal Model Essence Remains Unchanged

GPT-6 Astra has garnered attention through weight adjustments, but controlled experiments show that the core parameters of the model remain unchanged. Its performance improvement is due to weight optimization rather than intelligence enhancement. The model has a context length of 128k tokens and fine-tuning parameters of approximately 175B. In the MMLU benchmark test, it achieves a score of 82.5, with inference speed 3.5 times faster than the base version. It is currently only provided as a closed-source API for commercial services and is not open-sourced.