AI startup Inception has released the Mercury 3 diffusion language model, which claims to outperform GPT-6 Luna and Claude Haiku 5.5 on multiple benchmarks while operating at a fraction of the wait time. Rather than generating text sequentially word by word, the diffusion-based model sketches a rough version of an entire answer and refines it across multiple steps. In internal benchmarks, Inception reports that Mercury 3 generates over 1,000 tokens per second on Nvidia GPUs, significantly outpacing standard models that manage between 130 and 240 tokens per second.
In Inception's testing, Mercury 3 wins eight out of twelve evaluations against its rivals, showing particular strength in instruction following on IFBench and autonomous coding on SWE-Bench Atlas. However, competitors hold advantages in specific areas, with Claude Haiku 5.5 winning longer coding tests and GPT-6 Luna edging ahead in select Artificial Analysis evaluations. The model supports a one-million-token context window, features an OpenAI-compatible API, and has scaling capabilities that match larger parameter sizes with high processing speeds.
The architecture is targeted at real-time applications where cumulative latency impacts user experience, including autonomous AI agents, customer support platforms, smart glasses, and educational tutoring. Early-access partners report substantial performance gains, such as reduced response bottlenecks for wearable search and faster hint delivery for students. Launch pricing is temporarily halved to $0.10 for input and $0.35 for output tokens per million, though Inception has not specified the duration of the discount.
Why it matters: Mercury 3 demonstrates that diffusion models can scale effectively to rival traditional autoregressive language models while deliv