Inception says Mercury 3 beats GPT-6 Luna and Claude Haiku 5.5 at up to 8x the speed

Inception says Mercury 3 beats GPT-6 Luna and Claude Haiku 5.5 at up to 8x the speed

Diffusion language model startup Inception AI Inc. today released Mercury 3, its new flagship model.

In the company’s own benchmarks, Mercury 3 beat OpenAI Group PBC’s GPT-6 Luna and Anthropic PBC’s Claude Haiku 5.5 on eight of 12 tests at four to eight times their speed.

The release takes on the main doubt about Inception’s approach. Co-founder and Chief Executive Stefano Ermon said the question the company has heard most since Mercury 1 “is whether diffusion can scale.” As he sees it, the limits come down to “compute and data, not fundamental algorithmic barriers.”

GPT-6 Luna and Haiku 5.5 write their answers the conventional way, one token after another. A Mercury model drafts the whole response up front, then refines every position in parallel over a few passes. Inception clocked Mercury 3 at 1,068 tokens per second on Nvidia Corp. graphics processing units, with Haiku 5.5 reaching about 240 in the same comparison and GPT-6 Luna about 130.

Inception has not disclosed a parameter count for Mercury 3, saying only that it is more than 40 times the size of Mercury 1. The context window grows to 1 million tokens, up from the 260,000 that Mercury 2.5 handled. Ermon said training on Mercury 3’s successor is already under way.

Across the benchmark results, Mercury 3’s widest lead over GPT-6 Luna came in coding, where it scored 51.1 on SWE-Bench Atlas against 33.3 for the OpenAI model and 32.3 for Haiku 5.5. The new model trailed Haiku 5.5 on DeepSWE, a second coding benchmark, by 46.3 to 51.6, and on one long-context test. GPT-6 Luna won two more by less than a point each. No outside group has published its own results on the model yet.

Early customers are putting Mercury 3 into products where users notice every delay. 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. At tutoring app RevisionDojo, co-founder and Chief Technology Officer Michael Tong said hints and marked answers now come back in about a second.

Mercury 3 lists at 20 cents per million input tokens and 70 cents per million output tokens, with cached input 90% cheaper. A launch promotion halves both rates for now. GPT-6 Luna and Haiku 5.5 each list at 10 cents and 50 cents for typical prompts. Developers can reach the new model on Inception’s own platform or through OpenRouter Inc., Baseten Labs Inc. and Vercel Inc., and software already written for OpenAI’s application programming interface needs only minor changes to call it.

Inception is a venture capital-backed startup that Ermon founded in 2024 with fellow computer science professors Aditya Grover and Volodymyr Kuleshov. Microsoft Corp.’s M12 venture arm was among the backers when Menlo Ventures led a $50 million round in the company last November.

Microsoft was in talks to acquire Inception as of May, according to a Reuters report that said the startup wanted more than $1 billion. Space Exploration Technologies Corp. had also approached the company, Reuters said. No deal with either suitor has been announced.

Image: Inception AI