**dLLM architecture, which relies on iterative denoising rather than next-token autoregressive prediction, has delivered notable inference speed gains—often exceeding 1,000 tokens per second in models like Inception Labs’ Mercury 2 and Google’s DiffusionGemma—but has not yet produced a system that tops the Chatbot Arena leaderboard.** Recent scaling milestones, including LLaDA2.0 reaching 100B parameters and matching mid-tier autoregressive models on select math and coding benchmarks, demonstrate progress in parallel generation and error correction. However, frontier autoregressive systems from Anthropic, OpenAI, and Google continue to lead in general reasoning and overall Arena scores as of late September 2026. With only three months remaining before the December 31, 2026 resolution date, the absence of any dLLM surpassing the current top models underpins the market’s strong 85.4% implied probability for “No.” Key near-term catalysts include potential updates to Gemini Diffusion or new open-weight releases that could alter Arena positioning.