Inception launches Mercury 3, a diffusion LLM it claims tops GPT-6 Luna, Claude Haiku 5.5, and GLM-5.3 Flash on agentic benchmarks while being 4x to 8x faster

AI responses become nearly instantaneous and cheaper to generate.

Inception has launched Mercury 3, a diffusion-based large language model (LLM) that claims to outperform GPT-6 Luna, Claude Haiku 5.5, and GLM-5.3 Flash on agentic benchmarks while operating 4x to 8x faster. The model utilizes parallelized inference, generating multiple tokens simultaneously instead of sequentially, which significantly boosts speed and efficiency compared to traditional autoregressive LLMs. This approach leverages GPU parallel processing capabilities more effectively. Inception also reported that Mercury 3 agents achieved a 16% speed increase through recursive self-improvement. The model is priced at $0.10 per million input tokens and $0.35 per million output tokens, with a temporary 50% launch discount. This development challenges the prevailing industry strategy of scaling by increasing LLM size and relying on hardware advancements.

Mercury 3's core innovation lies in applying diffusion model technology, typically used for image and video generation, to LLMs. This allows for parallelized token generation and refinement, a stark contrast to the sequential, autoregressive process of conventional LLMs. This architectural shift not only accelerates response times dramatically but also promises greater efficiency in GPU utilization, potentially lowering operational costs for AI services.

This breakthrough could redefine the economics of AI development and deployment, shifting the focus from simply building larger models to optimizing underlying generation processes. The success of diffusion LLMs like Mercury 3 and Google DeepMind's DiffusionGemma may signal a broader industry pivot, challenging the current paradigm that equates AI advancement solely with increased model size and compute power. The long-term impact will depend on whether this technology scales to rival frontier models or remains a specialized solution for real-time applications.