Inception 称 Mercury 3 在多项基准上胜过 GPT-6 Luna 与 Claude Haiku 5.5 且快 4 到 8 倍

Inception says its new language model, Mercury 3, beats GPT-6 Luna and Claude Haiku 5.5 on a range of benchmarks while running four to eight times faster.

In its announcement, the company calls Mercury 3 the most capable diffusion language model released so far.

The speed comes from how the model is built. Models like GPT or Claude write one word at a time, so each new word has to wait for the one before it. Diffusion models are best known from image generation. They first sketch a rough version of the whole answer and then refine it over several passes, generating many words at once. Google's technical report on DiffusionGemma explains the basic idea.

Mercury 3's lead is narrower than the announcement suggests

Inception's own benchmark table tells a more mixed story. Mercury 3 wins eight of 12 tests, with all three models running at medium settings. The gap is widest in instruction following. On IFBench, Mercury 3 scores 76.84 points, 15 more than GPT-6 Luna. It also leads clearly in autonomous coding on SWE-Bench Atlas with 51.08 points, while its rivals reach 33.33 and 32.26.

Claude Haiku 5.5 wins the longer coding test DeepSWE. It's also better at finding information in long texts. GPT-6 Luna edges ahead in two tests from Artificial Analysis. Inception's short overview is more cautious, describing Mercury 3's quality as on par or better.

According to Inception, Mercury 3 generates 1,068 tokens per second on Nvidia GPUs. In the company's comparison, Claude Haiku 5.5 manages about 240 and GPT-6 Luna about 130. No independent tests exist yet.

Inception says Mercury 3 has more than 40 times as many parameters as Mercury 1 but runs just as fast. Performance has also tripled since the small Mercury 1 mini. Inception sees this as a sign that diffusion models reliably improve as they scale, just like standard language models. The exact parameter count remains undisclosed.

Mercury 3 targets apps where small delays add up

Inception is pitching Mercury 3 mainly for AI agents, search, customer support, and tutoring. AI agents complete tasks in many separate steps and wait on the model at each one. Those small delays pile up fast.

Early-access partners share their results in the announcement, and most of them also focus on speed. Octen AI builds search features for smart glasses. "On smart glasses, an answer that takes five seconds feels broken," says founder Kuan Zou. He says the search itself took only about 80 milliseconds, and the language model was the bottleneck. With Mercury 3, the median response arrived after 0.9 seconds. The full process took 2.2 seconds.

Education company General Learning says it now delivers hints and corrections to students in about one second. Customer service provider Lorikeet reports faster response times along with better results. At TinyFish AI, co-founder Shuhao Zhang says Mercury 3 was given 50 navigation tasks on real websites. It solved 25, beating GPT-6.1-Sol. Zhang doesn't give a number for the OpenAI model.

Mercury 3's low launch pricing may not last

Regular pricing is $0.20 per million input tokens and $0.70 per million output tokens. Inception is cutting prices in half at launch. It hasn't said when the discount ends. After that, Mercury 3 would cost more for output than GPT-6 Luna and Claude Haiku 5.5, at least for shorter inputs, where both charge $0.50. For very long prompts, though, both rivals charge a lot more. Mercury 3 can handle up to 1 million tokens at once.

The model is available on Inception's own platform as well as OpenRouter, Baseten, and Vercel. It uses the same API as OpenAI's models, so Inception says developers can switch with little effort.

In February, Inception released Mercury 2, which it called the first diffusion model to reason through intermediate steps before answering. Before that, the startup raised $50 million. Backers include NVentures, M12, and Snowflake Ventures, the venture arms of Nvidia, Microsoft, and Snowflake.

Google is working on the same approach. Its Gemini Diffusion demo was followed in June by the open model DiffusionGemma. Google says it runs up to four times faster than the standard Gemma 4 but delivers slightly weaker results.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.