Inception's Mercury 3 beats GPT-6 Luna, Haiku 5.5 on 8 of 12 tests

Inception has been trying to convince the market that diffusion language models—which draft an entire response up front and then refine every position in parallel, rather than writing one token after another—can scale. CEO Stefano Ermon said the question the company has heard most since Mercury 1 "is whether diffusion can scale," and he argues the limits come down to "compute and data, not fundamental algorithmic barriers." Mercury 3 is the company's attempt to prove that point, and its benchmark claims, though self-reported, are framed as a direct challenge to the conventional approach used by OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5.

The company has not disclosed a parameter count for Mercury 3, saying only that it is more than 40 times the size of Mercury 1. Early customers are already putting the model into products where users notice every delay: Octen AI's Kuan Zou said that on smart glasses "an answer that takes five seconds feels broken," and Octen's tests measured a median generation time of 0.9 seconds. At tutoring app RevisionDojo, CTO Michael Tong said hints and marked answers now come back in about a second.

Inception is a venture capital-backed startup founded in 2024 by Ermon with fellow computer science professors Aditya Grover and Volodymyr Kuleshov. Microsoft's M12 venture arm was among the backers when Menlo Ventures led a $50 million round last November. According to a Reuters report, Microsoft was in talks to acquire Inception as of May, with the startup seeking more than $1 billion, and SpaceX had also approached the company; no deal with either suitor has been announced.