Mercury 2 hits 1,000 tok/sec

Mercury 2 is one of the clearest shots yet at the autoregressive status quo: if the quality holds up in real workloads, speed stops being a UX tax and starts becoming a product advantage.

• –The real story is not just headline throughput, but that diffusion-based generation changes the latency curve for agent loops, coding copilots, voice systems, and RAG pipelines.
• –Inception is positioning Mercury 2 as a drop-in API replacement, which lowers adoption friction for teams already built around OpenAI-style tooling.
• –The model looks strongest for structured output, search, real-time interaction, and coding assistance, where sub-second responsiveness matters more than squeezing out every last point of frontier reasoning quality.
• –This launch also puts pressure on mainstream model vendors to show better speed-quality tradeoffs, not just bigger benchmark numbers.
• –Outside commentary already frames Mercury 2 as part of a likely hybrid future, where diffusion models handle fast draft generation and slower autoregressive models handle high-stakes refinement.