Inception's Mercury 3 Undercuts the Speed-Optimized Frontier at 1068 Tokens per Second

Inception Labs' Mercury 3 model, launched in October 2026, achieves 1068 tokens per second using diffusion-based architecture, significantly outperforming existing frontier models like GPT-6 Luna and Claude Haiku 5.5 in speed and cost. The model's pricing is aggressively lower, with a 90% discount on cached tokens, disrupting the market for high-throughput AI agent infrastructure by decoupling speed and cost from compute expenses.

This Wire brief sits within Fusion42's coverage of AI Frontier Models and AI Infrastructure.