Inception Unveils Mercury 3, Claims Speed and Performance Edge Over Competitors
Inception has introduced its latest language model, Mercury 3, asserting that it surpasses both GPT-6 Luna and Claude Haiku 5.5 in various benchmarks while operating four to eight times faster.
The company describes Mercury 3 as the most advanced diffusion language model to date. This speed advantage stems from its architectural design, which allows it to generate multiple words simultaneously, unlike traditional models like GPT or Claude that produce one word at a time.
Inception’s benchmark results present a more nuanced picture. While Mercury 3 leads in eight out of twelve tests, the competition performs well in specific areas. For instance, Mercury 3 scores 76.84 points on the IFBench for instruction following, outperforming GPT-6 Luna by 15 points. It also excels in autonomous coding on SWE-Bench Atlas with a score of 51.08, compared to its rivals’ scores of 33.33 and 32.26.
However, Claude Haiku 5.5 excels in longer coding tests such as DeepSWE and demonstrates better information retrieval from lengthy texts. GPT-6 Luna also outperforms in two tests from Artificial Analysis. Inception’s overview of Mercury 3 is cautious, indicating that its quality is comparable or superior.
According to Inception, Mercury 3 can generate 1,068 tokens per second on Nvidia GPUs, significantly faster than Claude Haiku 5.5 at approximately 240 tokens and GPT-6 Luna at about 130 tokens. No independent verification of these claims has been conducted yet.
Inception states that Mercury 3 has over 40 times the parameters of its predecessor, Mercury 1, while maintaining similar speed. The performance has reportedly tripled since the smaller Mercury 1 mini, suggesting that diffusion models improve reliably with scaling, akin to standard language models. However, the exact parameter count has not been disclosed.
Mercury 3 is primarily targeted at applications where minor delays can accumulate, such as AI agents, search, customer support, and tutoring. Early-access partners have shared their experiences, emphasizing the model’s speed. For example, Octen AI, which develops search features for smart glasses, noted that a five-second delay feels excessive. With Mercury 3, the median response time dropped to 0.9 seconds, with the entire process taking 2.2 seconds.
General Learning, an education company, reported that it now provides hints and corrections to students in about one second, while customer service provider Lorikeet has experienced faster response times and improved outcomes. TinyFish AI’s co-founder, Shuhao Zhang, mentioned that Mercury 3 completed 25 out of 50 navigation tasks on real websites, outperforming GPT-6.1-Sol.
Inception has launched Mercury 3 at a promotional price of $0.20 per million input tokens and $0.70 per million output tokens, halving the regular rates. The company has not disclosed when this discount will end. After the promotional period, Mercury 3 may become more expensive for output than GPT-6 Luna and Claude Haiku 5.5, particularly for shorter inputs, where both charge $0.50. However, for very long prompts, both competitors charge significantly more. Mercury 3 can accommodate up to 1 million tokens at once.
The model is accessible on Inception’s platform as well as through OpenRouter, Baseten, and Vercel, utilizing the same API as OpenAI’s models, allowing for easy transitions for developers.
In February, Inception released Mercury 2, which it described as the first diffusion model capable of reasoning through intermediate steps before generating answers. The startup has also raised $50 million in funding from investors including NVentures, M12, and Snowflake Ventures, the venture arms of Nvidia, Microsoft, and Snowflake.
Google is exploring similar approaches, having demonstrated its Gemini Diffusion model and later releasing the open model DiffusionGemma, which reportedly operates up to four times faster than the standard Gemma 4, albeit with slightly lower performance.
Source: The Decoder