A widening gap separates million-token inference from RL post-training, which remains at 256K tokens or below. The gap matters for AI agents, whose observations, tool outputs, documents, and decisions accumulate over long trajectories. Unlike inference, GRPO scores and backpropag…