Everyone’s debating Claude vs. GPT vs. Gemini.

Meanwhile, Inception Labs just quietly rewrote the rules of the game entirely.

Three professors from Stanford, Cornell, and UCLA just dropped Mercury 2 — the world’s first reasoning diffusion language model. And the numbers are kind of wild.

1,009 tokens per second.

Compare that to Claude Haiku 4.5 with reasoning at ~89 tokens/sec. Or GPT-5 Mini at ~71 tokens/sec. That’s not an incremental improvement. That’s a different category.

Here’s what makes this more than a speed flex:

The architecture is fundamentally different. Every major LLM you use today generates text one token at a time, left to right, sequentially. Mercury 2 generates entire blocks in parallel, refining them like an editor reworking a whole draft — not a typist adding one word at a time.

The price gap is just as striking. Mercury 2 runs at $0.25 per million input tokens and $0.75 per million output. That’s 4x cheaper than Gemini Flash on output and ~6x cheaper than Claude Haiku on output.

They’re not at the frontier yet. On reasoning benchmarks, Mercury 2 is competitive with Haiku and GPT-5 Mini — not Claude Opus or Gemini Pro. But that’s not the point.