1,068 tokens per second. That’s the number Inception wants you to remember from the Mercury 3 launch on October 19, 2026. It’s a bold claim. And in this Mercury 3 AI review, the honest answer is that the real story sits somewhere below it.
What makes Mercury 3 different
Most AI models write one token at a time. Left to right. No stopping, no revising, just the next word and then the next one after that, until the answer finally ends.
Mercury 3 is a diffusion language model, which flips that process around. It starts with a rough draft of the whole answer, then refines all of it at once over a few passes. Think of an image generator pulling a picture out of static. Same idea, applied to text. Because the model writes in parallel, long answers don’t have to wait for thousands of single steps.
Speed and price: read the claims carefully
So is it really that fast? Inception says Mercury 3 generates 1,068 tokens per second on NVIDIA GPUs, and its own benchmark table puts the model ahead of GPT-6 Luna and Claude Haiku 5.5 on 10 of 12 tests, all run by Inception itself at medium effort. Those are company claims, not verified results. The one outside measurement tells a cooler story: OpenRouter recorded a median of 160 to 254 tokens per second during launch week. Still quick. Just not 1,068.
The price is easier to check. List price is $0.20 per million input tokens and $0.70 per million output tokens, with a 1M-token context window for long documents and long chats. At launch, Inception cut that in half: $0.10 in, $0.35 out. New API keys also get 100 million free tokens, which gives you a lot of testing before you pay anything.
How to try it
You don’t need new tools to test it. The API is OpenAI compatible, and the model name is mercury-3, so code that already talks to OpenAI-style endpoints mostly needs a new address and a new key. You can reach it four ways: directly through the Inception API, or through OpenRouter, Baseten and Vercel AI Gateway. Puter.js offers another route for quick tests, with no API key at all. One warning if you’re on an older version: Mercury 2 and Mercury 2.5 are closed to new users, and existing customers keep access only until November 15, 2026.
Who should care
Vercel points Mercury 3 at work where waiting hurts: browser agents, real-time search, interactive education and customer support. That makes sense. If your app makes twelve model calls in a row, small delays stack up fast, and a model that writes in parallel earns its keep. If you ask one short question at a time, though? You’ll barely notice any of this.
The honest limits
Now the fine print. Mercury 3 takes text only, so there’s no image or audio input. Its weights are closed and Inception hasn’t shared the model’s size, which means no self-hosting and no peeking inside. No independent benchmark has scored it yet. And that launch price is a promotion with no announced end date, so smart budgeting uses the list price, not the discount.
Different by design, fast on paper, cheap while the launch discount lasts. My honest take: the idea is strong enough to deserve a test on your own work, not strong enough to trust on a launch announcement alone. Try the free tokens. See what happens.
Sources
- https://www.eesel.ai/blog/inception-mercury-3
- https://vercel.com/changelog/mercury-3-now-available-on-ai-gateway
- https://developer.puter.com/tutorials/free-unlimited-inception-mercury-api/