Inception's Mercury 3 outruns GPT-6 Luna at 1,000+ tokens a second
The diffusion language model matches GPT-6 Luna and Claude Haiku 5.5 on benchmarks while running 4 to 8 times faster, with a 1 million token context window and launch pricing cut to $0.10 per million input tokens.
Inception released Mercury 3 on 19 October, the newest version of its diffusion language model, a generative model that produces text by refining a whole draft at once rather than predicting one token after another. Inception calls it the most capable diffusion LLM released so far.
Mercury 3 runs at 1,068 tokens a second on Nvidia GPUs, has a 1 million token context window (up from 260,000 in Mercury 2.5), and is over 40 times the size of the original Mercury while matching its speed. Inception says it beats GPT-6 Luna and Claude Haiku 5.5 on standard benchmarks while running 4 to 8 times faster. It’s live through Inception’s own API, OpenRouter, Baseten and Vercel’s AI Gateway, with OpenAI-compatible endpoints. Launch pricing is $0.10 per million input tokens and $0.35 per million output, half the regular rate, with a free tier of 100 million tokens.
Why it matters: diffusion models have mainly been a speed trade-off, fast but behind on quality. Inception’s own figures claim that gap has closed without losing the speed, which matters for anything where latency compounds: voice pipelines, search agents, coding subagents making many small calls.
The caveat: every comparison here is Inception’s own, there’s no independent benchmark yet, and the launch price is explicitly temporary.