Inception's Mercury 3 hits 1,068 tokens per second

Inception's pitch rests on a design choice rather than raw scale. Mercury 3 is a diffusion model, which drafts many tokens at once and refines them, in contrast to autoregressive models that build answers one token at a time. The company says the model has more than 40 times Mercury 1's total parameters while maintaining similar speeds — an attempt to show the approach can grow without giving up its core advantage.

The launch lands in a market where speed is already a selling point. OpenAI announced GPT-6 Astra Ultrafast on September 29, advertising up to eight times faster token generation in Codex and up to six times faster generation in the API, and its January 14 agreement with Cerebras calls for 750 MW of compute capacity designed for fast responses. Inception is betting a different model design can give it an edge against that hardware-and-pricing push.

The practical question for buyers is narrower than the headline speed figure. OpenAI's own Ultrafast documentation warns that time spent opening connections can eat into gains from faster generation, and recommends keeping a WebSocket connection open for agents that call tools repeatedly. Faster text generation can still leave users waiting on the rest of a workflow, which is why Inception points to live tutors, voice support and browser agents rather than hard reasoning as Mercury 3's clearest opening.