Inception's Mercury 3 diffusion model claims 1,068 tokens a second, and it's half price at launch

Inception, which builds diffusion-based language models, launched Mercury 3 on October 19 and calls it the most capable diffusion LLM released so far. It says the model runs at 1,068 tokens per second on NVIDIA GPUs and matches or beats GPT-6 Luna and Claude Haiku 5.5 on its benchmarks while running 4 to 8 times faster. Those numbers come from Inception.

Context goes up to 1M tokens, from 260K on Mercury 2.5, with up to 131,072 output tokens and four reasoning effort levels: instant, low, medium and high. Regular pricing is $0.20 input and $0.70 output per million tokens. The launch price is half that, and cached input is $0.01. The pages I read don't say when the discount ends.

Mercury 3 is on Inception's own API and on OpenRouter, Baseten and Vercel, and the API is OpenAI-compatible. The launch post offers 100 million free API tokens to start. Early partners quoted in the post use it for browser agents, tutoring, search and support, and TinyFish says it completed 50% of 50 live Mind2Web navigation tasks in its own harness.