Mercury 3 hits 1,068 tokens/sec. Don't price on the discount.

Inception Labs shipped Mercury 3 today, the newest version of its diffusion LLM. It writes 1,068 tokens per second on Nvidia GPUs, the context window grows to 1 million tokens, and list price is $0.20 input / $0.70 output per million tokens. There is a 50% launch discount. The page does not say when it ends. That one missing date should shape how you build on it.

What actually happened

Per Inception's announcement:

• Speed: 1,068 tokens/sec on Nvidia GPUs. Inception claims 4–8x faster than comparable models.
• Context: 1M tokens, up from 260K on Mercury 2.5.
• Price: $0.20 / $0.70 per million tokens at list. With the launch discount: $0.10 / $0.35. Cached input gets 90% off.
• Quality: Inception says Mercury 3 matches or beats GPT-6 Luna and Claude Haiku 5.5 "across a range of benchmarks." It publishes no scores for that claim.
• Access: OpenAI-compatible API on Inception's platform, plus OpenRouter, Baseten and Vercel. New accounts get 100 million free tokens.

Diffusion models do not write one token at a time. They draft a whole block and refine it in parallel, which is where the speed comes from.

Why a 1,000-token-per-second LLM matters for your business

Speed changes what you can build, not just how fast it feels. At 1,000+ tokens per second, a full product description, support reply or SQL query comes back in well under a second. That makes some features possible that were too slow before: autocomplete in a form, a rewrite on every keystroke, an agent that runs ten small steps before the user notices.

But two habits keep this from going wrong:

1. Budget at list, not launch price. A discount with no end date is a marketing line, not a contract. If your margin only works at $0.35 output, your margin does not work.
2. Test quality on your own tasks. "Matches Haiku on a range of benchmarks" with no numbers is a claim. Run 50 of your real prompts through both and grade them blind.

Because Mercury 3 speaks the OpenAI API format, the test costs almost nothing: change a base URL and a model name. That is the right way to try any new model — behind a config flag, with an eval, and with an easy way back.

Sources: Inception Labs.