Inception Mercury 3: what it is and what it costs

Mercury 3 is a new language model from Inception that writes text by diffusion instead of one token at a time, and the company says it runs at more than 1,000 tokens per second. At launch it costs $0.10 per million input tokens and $0.35 per million output tokens, half the list price, and new API accounts get 100 million free tokens to try it.

Inception co-founder Stefano Ermon announced it on Monday. In his thread, @StefanoErmon called it "the most intelligent diffusion LLM ever released" and said it beats GPT-6 Luna and Claude Haiku 5.5 on quality at that speed. A few minutes later, co-founder @volokuleshov wrote that it is about three times smarter than Mercury 1 on standard benchmarks and runs at the same speed despite being more than 40 times larger. Early partners, he said, use it for browser agents and real-time tutoring.

What a diffusion LLM is

Nearly every chatbot you've used writes the way you type on a phone: one token, then the next, then the next. A text diffusion model drafts a whole block at once and refines many positions in parallel, the same basic idea image generators use to sharpen a picture out of noise. You give up some of the step-by-step character of normal models, and you get speed.

The doubt about diffusion for text has been whether it can get smart as well as fast. Inception's launch post says plainly that "text diffusion scales." Mercury 3 is the company's fourth model since Mercury 1 in June 2025, and this time the comparison it picks is with the small models from OpenAI and Anthropic.

What Mercury 3 can do, per Inception

• It runs at 1,068 tokens per second on NVIDIA GPUs, which Inception says is 4 to 8 times faster than comparable models
• The context window is 1 million tokens, up from 260,000 on Mercury 2.5
• @OpenRouter, which listed the model the same day, relayed Inception's claim that it beats GPT-6 Luna and Haiku 5.5 on IFBench, GPQA Diamond and SWE-Bench Atlas
• In Inception's own web-agent harness, it finished 50% of 50 live Mind2Web navigation tasks, more than GPT-6.1-Sol managed in the same test
• It speaks the OpenAI API format, so existing code can point at it without a rewrite

Is Mercury 3 free, and where can you use it?

Partly. You can chat with it through Inception's Mercury Chat, and the API platform comes with 100 million free tokens. After that you pay per token. The list price is $0.20 per million input and $0.70 per million output, with 50% off at launch and cached input at 90% off. Inception hasn't said how long the discount runs. Besides its own API, the model is on OpenRouter, Baseten and Vercel.

The launch post says nothing about open weights, so for now you can't download it and run it on your own hardware.

My read

For a lot of real jobs, speed matters more than a few benchmark points. A voice agent that pauses for two seconds before every answer sounds broken. A support bot that takes ten seconds to draft a reply loses the customer. Inception is selling into those jobs: its post names voice agents, customer support, search agents and high-volume classification.

Treat the benchmark claims as claims for now. Every number above comes from Inception or partners repeating Inception, and the launch post gives no head-to-head scores against GPT-6 Luna or Haiku 5.5, only the statement that Mercury 3 wins. I do like that they published the Mind2Web number, but read it both ways: half the tasks finished means half didn't. Independent tests on OpenRouter should show up within days, and those are worth waiting for before you move anything important.

The price and the free tokens make it cheap to test, though. If you already have something running on a small model, swapping the endpoint and running the same 50 real requests through both would take an afternoon.

What this means for a business

Small, fast models are getting good enough for the boring, high-volume work that eats a team's day: sorting inbound emails, tagging support tickets, pulling fields out of invoices, answering the same ten customer questions. At a launch price of 35 cents per million output tokens, cost isn't what holds most owners back anymore. The harder part is knowing which of your tasks are repeatable enough to hand off. New Face Design's free process audit starts there: we map where your hours go and point out the few jobs a model like this could take off your plate.