On Oct. 19, Inception launched Inception Mercury 3, which it calls the most capable diffusion-based large language model released to date. The model has a 1M-token context window and is available now through Inception's API platform and three partner services.
What Mercury 3 is
In short, a diffusion LLM refines a whole draft of its answer in parallel, instead of writing one token at a time like most chat models.
Inception says Mercury 3 is more than 40x larger than Mercury 1 in total parameters, and that its context window grew from 260K tokens in Mercury 2.5 to 1M.
What Inception claims
By Inception's own numbers, Mercury 3 matches or beats GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks. It does so while running 4x to 8x faster, at more than 1,000 tokens per second. Inception's measured figure is 1,068 tokens per second on NVIDIA GPUs.
Inception also says benchmark intelligence rose 3x from Mercury 1 (mini) to Mercury 3. It presents that as evidence that scaling laws seen in conventional models also hold for diffusion.
Price and where to get it
| Item | Price per million tokens |
|---|---|
| Input, list | $0.20 |
| Output, list | $0.70 |
| Input, launch price | $0.10 |
| Output, launch price | $0.35 |
Prices are as of 9:55 PM ET Oct. 19. The launch price is 50% off. Inception's post gives no end date for the discount. Cached tokens cost 90% less than the input price.
Mercury 3 runs on the Inception API Platform and through OpenRouter, Baseten and Vercel. It is OpenAI API compatible, and Inception offers 100 million free tokens to try the API.
Who it is for
Inception points to latency-sensitive uses. Early-access partners run it in browser-use agents, interactive education, real-time search agents and customer support. Inception also lists Mercury Voice and Mercury Decide, both part of what it calls a real-time product portfolio.
Inception says agents wait on the model at every step, so small delays add up. Voice and chat support, and search that feels slow after a few seconds, are where that shows first.
What to check first
Start with the free tokens and run your own workload, with your own prompts and your own pass-fail checks. Partner quotes in the launch post are the partners' own claims, so judge quality on your own prompts.
Budget at the list price, since the discount has no stated end date. If you wire a new model into an app, our guide to rotating AI API keys and setting spending limits covers keeping the bill in check. For other developer AI launches, see Cloudflare's Clairic extraction model.
