Our reference case puts Grok 4.8 at 58 days after 4.7. Reinforcement learning, the training that improves a model using generated responses, remains the slower stage.

Flight 14 reached orbit on September 28 and deployed all 26 Starlink V3 satellites. Both stages then splashed down at sea, as planned, so this one flight carries the full construction cost of Booster 21 and Ship 41.

Our model puts the gap between Grok 4.7 and 4.8 at 52–99 days across the middle 80% of simulated outcomes. At our assumed speedup, an in-house reinforcement-learning stack would bring the reference release date forward by 26 days by shortening the work needed to improve the model’s responses before launch.