The AI That Writes the Whole Answer at Once

Almost every AI chatbot you have used writes the same way you would type a text message: one word at a time, left to right, with each new piece waiting on the one before it. That is why even the best assistants can feel a little slow when you ask for something long. Last Monday, October 19, a company called Inception released a model that works differently, and it is worth paying attention to.

The model is called Mercury 3, and it is what researchers call a diffusion language model. Instead of building a reply word by word, it sketches out the whole answer up front and then cleans up every part of it at the same time over a few quick passes. If you have ever watched an AI image generator turn a blurry mess into a sharp picture, it is the same basic idea, applied to text.

The obvious result is speed. Inception measured Mercury 3 at 1,068 tokens per second on NVIDIA graphics cards. For comparison, SiliconANGLE reported that in the same test, Anthropic's Claude Haiku 5.5 reached about 240 and OpenAI's GPT-6 Luna about 130. In plain terms, Mercury 3 can finish a long answer before the others are a quarter of the way through.

Speed alone is not new for this approach, though. The big question since Inception launched its first model in June 2025 was whether diffusion models could get smart enough to matter. According to the company, Mercury 3 is more than 40 times larger than Mercury 1, and in its own benchmarks it beat GPT-6 Luna and Haiku 5.5 on eight of twelve tests. Its biggest lead came in coding. It also now handles a context window of one million tokens, up from 260,000 in the previous version, so it can work with very long documents in a single go.

There are fair caveats. Those benchmark wins come from Inception itself, and it did not win everything. It trailed Haiku 5.5 on one coding test and on a long context test. One independent developer who spent about $56 testing it across more than 62,000 calls summed it up as genuinely fast and genuinely good, but not automatically the cheapest or most accurate choice for every job. It also costs a bit more per token than the two models it is compared against, though a launch discount cuts the price in half for now.

So why should a small business care? Because a lot of the AI tools showing up in websites and apps today are things people sit and wait on: a support chat answering a customer, a booking assistant checking several details at once, a search tool pulling up the right page. When those replies arrive almost instantly, the experience stops feeling like talking to a robot and starts feeling like a helpful person who already knew the answer.

Mercury 3 is available now through Inception's own platform and services like OpenRouter, Baseten and Vercel, and it works with code already written for OpenAI's tools. We will be keeping an eye on how it holds up in real projects. If faster, smoother AI features sound useful for your site or app, that is exactly the kind of thing we like to build.