Inception launches Mercury 3, a fast diffusion model at $0.20 per million input tokens

Inception has released Mercury 3, a language model that writes with diffusion, a different method from the one most chat models use. The company says it beats GPT-6 Luna and Claude Haiku 5.5 on a range of benchmarks while running 4 to 8 times faster, at 1,068 tokens a second on NVIDIA GPUs, with a 1M-token context window. List price is $0.20 per million input tokens and $0.70 per million output, half that during the launch offer, and cached input is 90% off. It works with the OpenAI API format, is on Inception's own API (with 100 million free tokens to try), OpenRouter, Baseten and Vercel, and early users include a customer support company and a tutoring app. The benchmark claims are Inception's own. On Hacker News (218 points, 84 comments), 10 of the 26 top-level commenters tried it. Eight found it fast and capable, for coding, a small Mac app built end to end and test drawings, though two saw a slow first response or 300 to 400 tokens a second rather than 1,000. Two were less impressed: one says DeepSeek V4.1 Flash on OpenRouter beat it on speed and quality, and one found it slower than GPT-6.1 Sol on a maths puzzle. Four more are positive without testing, two doubt the comparisons (one questions how the charts are scaled), seven ask questions or want open weights, image input or a coding plan, and three are other posts, including the co-founder's launch note. Testers lean positive, but there is no consensus on whether it beats its rivals. Sorted take: For chat, voice or support flows where waiting is the problem, this is cheap to test with the free tokens; run it on your own prompts against your current fast model, and price it at the full rate, not the launch discount.