U.S. AI startup Inception announced its diffusion-based language model Mercury 3 in October 2026. According to the company’s comparison, it generates 1,068 tokens per second on Nvidia GPUs—four to eight times faster than GPT-6 Luna and Claude Haiku 5.5. Inception also claims Mercury 3 outperformed the others in eight of 12 benchmarks, offering an example of how the AI race is expanding from “intelligence” to “latency.”
What lies behind a model that doesn’t write one word at a time
Mainstream language models such as GPT and Claude generate text one word at a time. The next word can’t be produced until the previous one is determined, much like waiting in a checkout line. No matter how fast the computer is, this “one-at-a-time” structure remains unchanged.
Mercury 3 uses an approach called a diffusion model. It applies to text an idea that became widespread in image-generation AI: first, it creates a rough draft of the entire answer, then refines it over several rounds of processing. Inception says this gives it a speed advantage because it can handle many words at once. Google is pursuing research along similar lines. After demoing Gemini Diffusion, it released the open model DiffusionGemma in June. It is said to be up to four times faster than standard Gemma 4, though its results are somewhat worse.
According to Inception, Mercury 3 has more than 40 times as many parameters as the original Mercury 1, yet runs at the same speed. The company says it delivers three times the performance of the smaller Mercury 1 mini. Inception sees it as an important sign that diffusion models can improve steadily as they scale up.
Benchmark results aren’t a clean sweep
The headline figure in the announcement is “first place in eight of 12 benchmarks.” But the comparison was conducted by the company itself, with all three models running at medium settings.
The biggest gaps were in following instructions. Mercury 3 scored 76.84 on IFBench, about 15 points ahead of GPT-6 Luna. It also scored 51.08 on SWE-Bench Atlas, which measures autonomous coding, outpacing competitors at 33.33 and 32.26.
Claude Haiku 5.5, however, came out on top in DeepSWE, a longer coding challenge. It also led in tasks that involve finding information in longer texts. GPT-6 Luna had a slight edge in two Artificial Analysis benchmarks. That breakdown may explain why the company’s brief introduction cautiously describes the model’s quality as “on par with or better than” its rivals.
In Inception’s comparison, Mercury 3 generates 1,068 tokens per second, versus about 240 for Claude Haiku 5.5 and about 130 for GPT-6 Luna. Independent third-party verification has not yet been published. It’s sensible to treat the figures as a “promising claim” and test the model for yourself in your own use case.
When latency becomes a selling point
Inception cites AI agents, search, customer support, and educational assistance as key applications. Agents in particular break a task into multiple steps, waiting for the model’s response at each one. Even if each delay is small, the cumulative effect can make a big difference to how fast the system feels to users.
Early users have focused on speed, too. Kuan Zou, founder of Octen AI, which develops search features for smart glasses, said, “On smart glasses, an answer that takes five seconds feels broken.” The search itself took about 80 milliseconds, he said, with the language model the bottleneck. With Mercury 3, median response time fell to 0.9 seconds, and total processing time was 2.2 seconds.
Education company General Learning says it can now deliver hints and feedback to students in about one second. Customer service company Lorikeet reported improvements in results as well as faster responses. TinyFish AI, which automates website interactions, says it solved 25 of 50 navigation tasks on live websites, outperforming GPT-6.1-Sol. However, it did not provide specific figures for OpenAI’s model.
Look beyond speed when deciding whether to adopt it
Getting started is relatively easy. Mercury 3 is available through Inception’s own platform as well as OpenRouter, Baseten, and Vercel. Inception says switching from existing systems requires little effort because the model uses an API compatible with OpenAI’s. It can handle inputs of up to one million tokens.
At standard rates, input costs $0.20 per million tokens and output costs $0.70. At the time of the announcement, the model was available at half price, but the company did not say how long the discount would last. At standard rates, its output price is higher than that of GPT-6 Luna and Claude Haiku 5.5, both at $0.50, for short inputs. For very long inputs, however, the two competitors cost considerably more.
In February 2026, Inception released Mercury 2, which it described as the first diffusion model to include intermediate reasoning before responding. The company had previously raised $50 million, with backers including Nvidia, Microsoft, NVentures, M12, and Snowflake Ventures.
Ultimately, speed alone isn't enough to determine which model to choose. The decision depends on a combination of factors: quality, cost, how well it handles long texts, and how sensitive your service is to latency. For applications where real-time performance is critical, Mercury 3 looks like a candidate worth trying.
Comments