# Mercury 3: a diffusion language model that writes over 1,000 tokens a second

> Inception released Mercury 3 on Monday. Mercury 3 is a diffusion model: it drafts the whole answer at once and refines it in a few passes.

Published 2026-10-19T13:09:36-07:00 · News · Canonical: https://news.littlewolflabs.com/posts/2026-10-19-inception-mercury-3

[Inception](https://www.inceptionlabs.ai/blog/introducing-mercury-3) released Mercury 3 on Monday. Most chatbots write one word piece at a time. Mercury 3 is a [diffusion model](/posts/diffusion-llms-and-types-of-language-models): it drafts the whole answer at once and refines it in a few passes. [Inception](https://x.com/StefanoErmon/status/2112222911218291051) says it matches or beats OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 on its benchmarks while generating 1,068 tokens per second on Nvidia GPUs. It has a one-million-token context window, up from 260,000 for Mercury 2.5. List price is $0.20 per million input tokens and $0.70 per million output tokens, and it's half price at launch ($0.10 and $0.35). Cached input is 90% off. You can use it through Inception's own API, where new keys get 100 million free tokens, and through OpenRouter, Baseten and [Vercel's AI Gateway](https://vercel.com/changelog/mercury-3-now-available-on-ai-gateway). It works as a drop-in for code written for OpenAI's API.

*What this means for you:* speed is the selling point. It suits voice assistants, live chat, real-time search and browser agents, where waiting for a reply hurts the experience. All of the quality claims come from Inception itself, and nobody has independently checked them yet, so test it on your own prompts. If you use [Mercury 2 or 2.5](https://www.inceptionlabs.ai/models), take note: both are closed to new users, and existing customers with usage before Oct 19 keep access only until Nov 15, 2026, so plan your move.