Mercury 3 Diffusion Model Launch: Speed Record Set

Most AI models write as you do, one word at a time, left to right, with no going back. Now imagine a model that drafts a whole answer at once and then polishes every word in parallel. That is diffusion, and it just got a flagship moment. The Mercury 3 diffusion model launch on October 19, 2026 promises a staggering 1,068 tokens per second with a one-million-token context window. In this article, you will learn how Mercury 3 works, what its speed and price claims mean for you, and whether it is worth switching your stack for.

What the Mercury 3 Diffusion Model Launch Announced

Inception, a Redwood City startup, released Mercury 3 on October 19, 2026, calling it its most capable diffusion-based language model to date. Diffusion is a different way of writing text. Instead of predicting one token after another, the model sketches a rough draft of the whole response and then refines every position in a few parallel passes. Inception says Mercury 3 hits 1,068 tokens per second on NVIDIA GPUs, runs four to eight times faster than GPT-6 Luna and Claude Haiku 5.5, and carries a one-million-token context window, up from 260,000 in Mercury 2.5. The company also says it is more than 40 times larger than the original Mercury 1.

Why the Mercury 3 Diffusion Model Launch Matters for Speed

Speed is the whole point of this release. For voice assistants, coding copilots, real-time tutoring tools, and agent pipelines, slow replies kill the experience. If the claims hold, a 1,068 tokens per second model with a million-token context could pair deep reasoning with huge documents without the usual latency pain. Launch pricing is aggressive too. It is 0.10 dollars per million input tokens and 0.35 dollars per million output tokens, half of the 0.20 and 0.70 dollar list prices, with no announced end date for the discount. It is available through the Inception API, OpenRouter, Baseten, and Vercel, and the OpenAI-compatible API makes plugging it into an existing app simple.

Company Claims vs Independent Proof

Here is where a careful reader should slow down. The speed figures, the four- to eight-times faster comparison, and the benchmark wins over GPT-6 Luna and Claude Haiku 5.5 come from Inception itself. No independent benchmark has confirmed them yet, and the company has not said exactly which NVIDIA GPU produced the 1,068 figure. The claims may still be true, but treat them as a strong pitch, not a settled fact. Verify on your own workload before you commit. For context on how big model releases usually compare, our Mistral Large 4 explainer and our Meta Muse Wonder coverage show how launch claims are typically framed.

Who Should Try Mercury 3 First

If you build latency-sensitive products, this release is aimed directly at you. Voice bots, live search, tutoring apps, and customer support tools stand to gain the most, because a sub-second answer beats a slightly smarter slow one in real use. The million-token context window also suits anyone feeding a model huge files, long transcripts, or entire codebases in one pass. If you mostly need the deepest reasoning on hard problems, wait for independent tests. And if your team already uses OpenAI-style tooling, the compatible API makes a side-by-side trial cheap. Our OpenAI news explainer is a useful reference for comparing ecosystems.

Mercury 2 Shutdown and What Current Users Must Do

One detail is a deadline, not a promise. Mercury 2 and Mercury 2.5 are closed to new users, and existing non-enterprise access ends on November 15, 2026. If you run on those models, plan your migration now. Mercury 3 is meant as a drop-in upgrade, and the OpenAI-compatible API lowers the switching cost. Use the next few weeks to benchmark your real prompts on it and confirm the speed gains in your own app before the cutoff arrives.

Frequently Asked Questions

When did Mercury 3 launch? Inception released Mercury 3 on October 19, 2026.

What is a diffusion language model? It drafts a whole response at once and refines every word in parallel passes, instead of generating one token after another like standard models.

How fast and how big is Mercury 3? Inception claims 1,068 tokens per second on NVIDIA GPUs, four to eight times faster than GPT-6 Luna. And Claude Haiku 5.5, with a one-million-token context window and more than 40 times the size of Mercury 1. These are company claims.

How much does Mercury 3 cost? Launch pricing is 0.10 dollars per million input tokens and 0.35 dollars per million output tokens. A 50 percent discount on the list price with no announced end date.

The Mercury 3 diffusion model launch is the strongest signal. That diffusion can scale into a serious alternative to standard autoregressive models. The speed claims are exciting, the price is hard to ignore, and the migration clock is ticking for older Mercury users. Just remember the claims are company-reported for now, so test on your own data before you bet your product on them. Launch coverage worth reading includes the Unite AI report, Oossa coverage, and the Golden Era report.