Mercury 3: Diffusion at scale

Throughput is 1,068 tokens per second on NVIDIA GPUs, and the context window reaches 1M tokens, up from 260K for Mercury 2.5. List pricing is $0.20 per million input and $0.70 per million output tokens, and a launch discount halves those rates. The company says the model beats both Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna on three benchmarks, IFBench, GPQA Diamond, and SWE-Bench Atlas. The model produces output by refining a block of tokens in parallel, not by writing left to right. The company's first Mercury model, launched in 2025, already hit over 1,000 tokens per second on Nvidia H100s. The benchmarks are not independently verified, yet. Teams building voice agents, tutoring tools, or browser automation care most about latency and unit cost, and the vendor pitches this model at exactly those workloads. In a 50-task browser test, the company reports a 50% completion rate with token use similar to a frontier model at a lower unit price. These figures are self-reported, so a short evaluation on your own traces is the practical next step before switching.