Traders assign just 2.6% implied probability to a diffusion large language model (dLLM) claiming the top spot before 2027 because autoregressive frontier systems from OpenAI, Anthropic, Google, and xAI continue to lead on capability benchmarks while dLLMs remain secondary. Recent releases such as Google DeepMind’s DiffusionGemma (26B, June 2026) and Inception Labs’ Mercury 2 demonstrate strong inference throughput—often exceeding 1,000 tokens per second—and competitive results on code and reasoning tasks, yet they trail scaled autoregressive models in broad intelligence metrics. Ongoing scaling of transformer-based training runs, proven data and compute advantages, and the absence of any dLLM surpassing leaders on aggregate leaderboards underpin the near-certain consensus. A sudden breakthrough in dLLM scaling laws, major lab pivot, or regulatory slowdown on autoregressive development could still shift outcomes before year-end 2026.