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

Diffusion language model startup Inception AI Inc. has released Mercury 3, its new flagship model. In the company’s own benchmarks, Mercury 3 GPT-6 beat OpenAI Group PBC’s Luna and Anthropic PBC’s Claude Haiku by 5.5 in eight of 12 tests at four to eight times speed.

The publication clears up the main doubts about Inception’s approach. Co-founder and CEO Stefano Ermon said the question the company has heard most since Mercury 1 “is whether distribution is scalable.” In his view, the limits lie with “computing and data, not fundamental algorithmic barriers.”

GPT-6 Luna and Haiku 5.5 write their answers in the traditional way, one token at a time. A Mercury model designs the entire response in advance and then refines each position in parallel over a few passes. Inception clocked Mercury 3 on graphics processors from Nvidia Corp. at 1,068 tokens per second, with Haiku 5.5 reaching around 240 and GPT-6 Luna around 130 in the same comparison.

Inception did not disclose a parameter count for Mercury 3, only saying it is more than 40 times larger than Mercury 1. The context window grows to 1 million tokens, up from the 260,000 that Mercury 2.5 processed. Ermon said training for Mercury 3’s successor is already underway.

In terms of benchmark results, Mercury 3 achieved the biggest lead over GPT-6 Luna in encoding, where it achieved a score of 51.1 in the SWE bench Atlas, versus 33.3 for the OpenAI model and 32.3 for Haiku 5.5. The new model trailed Haiku 5.5 by 46.3 to 51.6 on DeepSWE, a second coding benchmark, and on a long-context test. GPT-6 Luna won two more by less than a point each. To date, no outside group has published its own results on the model.

Early customers are building Mercury 3 into products where users notice any lag. Kuan Zou, founder and CEO of real-time search startup Octen AI, said that on smart glasses, “an answer that takes five seconds feels broken.” The average 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 starred answers now come back in about a second.

Mercury 3 quotes 20 cents per million input tokens and 70 cents per million output tokens, with cached inputs being 90% cheaper. An introductory campaign halves both tariffs for the time being. GPT-6 Luna and Haiku 5.5 cost 10 cents and 50 cents respectively for typical prompts. Developers can achieve 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 requires only minor modifications to invoke it.

Inception is a venture capital-backed startup that Ermon founded in 2024 with 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.

According to a Reuters report, Microsoft was in talks to acquire Inception in May, with the startup seeking more than $1 billion. Space Exploration Technologies Corp. was also reported, according to Reuters. approached the company. No deal with any of the suitors has been announced.

Image: Inception AI