Mercury 3 Claims Speeds of Up to 8× Faster, but Adoption Should Follow Independent Verification
Inception Labs says its diffusion language model Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5 on selected benchmarks while running 4–8 times faster. Because the comparisons are vendor-reporte
Inception Labs launched its diffusion language model Mercury 3 on October 19, 2026, claiming that it outperformed GPT-6 Luna and Claude Haiku 5.5 on some benchmarks and was 4–8 times faster. If accurate, the claim could significantly reduce latency in tasks such as voice assistants, web agents, and real-time search that call models repeatedly. However, the comparison figures currently available come from the company’s own measurements, so the claimed performance advantage should be distinguished from perceived speed in real-world service environments.
Parallel Refinement Instead of Sequential Generation
Conventional autoregressive language models generate tokens sequentially, producing one token before continuing to the next. The diffusion approach described by Inception refines multiple token positions in parallel. This difference may be advantageous for quickly producing long answers, but actual response times are also affected by operating conditions such as inference settings, prompt length, tool calls, and network performance.
What the Published Figures Cover
The company reported 1,068 tokens per second on NVIDIA GPUs and a 1-million-token context window. At list prices, the cost is $0.20 per million input tokens and $0.70 per million output tokens; at launch, the company said these would be discounted to $0.10 and $0.35, respectively. The API is available through the Inception platform, OpenRouter, Baseten, and Vercel.
How to Use the Comparison Table as an Adoption Basis
Speed figures may be most valuable for multi-step agent tasks rather than short conversations. Quality comparisons, on the other hand, vary by task. According to reports, no public third-party evaluation of Mercury 3 is available yet, so when selecting a candidate model, it is reasonable to test accuracy, time to first token, total completion time, and cost using your own prompts, safety standards, and tool-use workflows.