Inception releases Mercury 3 with faster results in its own tests

Inception AI Inc. has released Mercury 3, its new flagship diffusion language model. In the company’s benchmarks, it beat OpenAI Group PBC’s GPT-6 Luna and Anthropic PBC’s Claude Haiku 5.5 on eight of 12 tests while running four to eight times faster.

The results address a question Inception has faced since Mercury 1. Co-founder and Chief Executive Stefano Ermon said people have asked “whether diffusion can scale.” He attributed the limits to “compute and data, not fundamental algorithmic barriers.”

Inception reports faster generation and mixed test scores

GPT-6 Luna and Haiku 5.5 generate answers one token at a time. Mercury drafts a full response, then refines its positions in parallel over several passes. Inception measured Mercury 3 at 1,068 tokens per second on Nvidia Corp. graphics processing units. In the same comparison, Haiku 5.5 reached about 240 and GPT-6 Luna about 130.

Mercury 3 scored 51.1 on the coding test SWE-Bench Atlas, compared with 33.3 for GPT-6 Luna and 32.3 for Haiku 5.5. It trailed Haiku 5.5 on another coding test, DeepSWE, by 46.3 to 51.6, and on one long-context test. GPT-6 Luna won two other tests by less than a point each. No outside group has published independent results for Mercury 3.

Inception has not disclosed Mercury 3’s parameter count. It says the model is more than 40 times the size of Mercury 1. Its context window is 1 million tokens, up from 260,000 in Mercury 2.5.

Customers test responses in real-time products

Early customers are using Mercury 3 where response time matters. Kuan Zou, founder and chief executive of real-time search startup Octen AI, said that on smart glasses “an answer that takes five seconds feels broken.” Median generation time in Octen’s tests was 0.9 seconds. Michael Tong, co-founder and chief technology officer of tutoring app RevisionDojo, said hints and marked answers now return in about a second.

Mercury 3 lists at 20 cents per million input tokens and 70 cents per million output tokens. Cached input costs 90% less, and a launch promotion currently halves both listed rates. Developers can access the model through Inception’s platform, OpenRouter Inc., Baseten Labs Inc. and Vercel Inc. Software built for OpenAI’s application programming interface needs only minor changes to call it.

Ermon founded the venture-backed company in 2024 with fellow computer science professors Aditya Grover and Volodymyr Kuleshov. Menlo Ventures led a $50 million funding round last November, with Microsoft Corp.’s M12 venture arm among the backers. Reuters reported that Microsoft was in acquisition talks with Inception as of May and that the startup wanted more than $1 billion. Reuters also said Space Exploration Technologies Corp. had approached Inception. No deal with either company has been announced.

Ermon said training on Mercury 3’s successor is already under way.