Mercury 3: a 1,068 tok/s diffusion LLM, live and OpenAI-compatible
Inception's new flagship runs 4-8x faster than rivals, launches at $0.10/M input, drops into any OpenAI stack, and handles 1M-token context today.
Copy markdown
A diffusion model that writes 1,068 tokens a second
Mercury 3 generates text in parallel instead of token-by-token, so Inception clocks it at 1,068 tok/s on standard NVIDIA GPUs — 4-8x faster than comparable autoregressive models. It's their flagship dLLM, over 40x the size of the first Mercury, and Inception says quality is on par with or better than GPT-6 Luna and Claude Haiku 5.5.
The launch price undercuts the small-model tier
List price is $0.20/M input and $0.70/M output, but launch pricing is 50% off: $0.10/M in, $0.35/M out, with cached input 90% cheaper. The API trial includes 100M free tokens, so you can A/B it against Haiku on your own workload before paying a cent.
Drop it into an OpenAI stack with no rewrites
It's OpenAI-API-compatible and already live on the Inception API, OpenRouter, Baseten, and Vercel — change the base URL and model name and you're running. The context window jumps to 1M tokens, up from 260K in Mercury 2.5.
Build this weekend: a sub-second agent or tutor
Inception reports a 0.9s median generation time and 2.2s end-to-end in a live search app, plus 50% completion on 50 live Mind2Web browser tasks — ahead of GPT-6.1-Sol. That speed makes real-time use cases practical: live translation, interactive tutoring with instant feedback, or a browser-use agent that doesn't stall.
Read the fine print before you switch
Every figure here is Inception's own — the blog claims wins over GPT-6 Luna and Haiku 5.5 but publishes no benchmark scores, and diffusion models can still trail autoregressive ones on long, exacting reasoning. Inception also shipped Mercury Voice and Mercury Decide alongside, so it's a family play. Trial it before you move production traffic.