# What Is Mercury 3? Inception’s Diffusion AI Model, Explained

URL: https://whatisagenticai.net/blog/what-is-mercury-3/
Published: 2026-10-19
Publisher: Little Tough LLC, maker of AI Agent Assistant: Agentic AI (https://apps.apple.com/us/app/ai-agent-assistant-agentic-ai/id6761501260)

Mercury 3 is Inception Labs’ new diffusion AI model, out Oct 19, 2026. How it writes 1,068 tokens a second, what it costs, and how to try it with free tokens.

Mercury 3 showed up in AI news on Monday with one number attached to almost every headline: 1,068 tokens a second. If you have never heard of the company behind it, or of a "diffusion" language model, that number does not tell you much.

So here is the short version. What Mercury 3 is, who makes it, why it works differently from ChatGPT or Claude, what it costs, and how you can try it today.

**TL;DR:** Mercury 3 is a diffusion-based large language model from Inception (often called Inception Labs), released October 19, 2026. Instead of writing one word at a time, it drafts many words at once and refines them, which lets it produce about 1,068 tokens per second on Nvidia GPUs. It costs $0.20 per million input tokens and $0.70 per million output tokens, half that during a launch promotion, and new API accounts get 100 million free tokens.

## What is Mercury 3?

Mercury 3 is the newest flagship model from Inception, a startup that builds diffusion large language models, or dLLMs. Inception announced it on its blog on October 19, 2026, and calls it "the most capable diffusion-based LLM released to date."

It is the third big step in the Mercury line. The first Mercury models came out in 2025, Mercury 2 followed in February 2026, and Mercury 2.5 arrived in September. Inception says Mercury 3 has more than 40 times the total parameters of Mercury 1, though it has not shared an exact parameter count.

People search for it in a few ways: "Mercury 3," "Inception Mercury 3," "Inception Labs Mercury 3," and "Mercury 3 AI." They all mean the same model. It has nothing to do with the boat engines or car amplifiers that share the name.

## What is a diffusion LLM, in plain English?

Most chatbots you know, like ChatGPT, Claude and Gemini, are autoregressive. That is a long word for a simple idea: they write one token (a small piece of a word) at a time, left to right, and each new token waits for the one before it.

A diffusion model works more like an editor with a rough draft. It starts with a fuzzy version of the whole answer, then cleans up many tokens at once over a few passes until the text is sharp. Image tools have worked this way for years. Inception’s bet is that the same trick can make text models much faster, and its new post says the scaling rules that helped regular models grow also hold for diffusion.

The payoff is speed. Because it is not stuck waiting on one token at a time, Mercury 3 can push out over 1,000 tokens a second. Inception claims that is 4 to 8 times faster than the models it compared against.

| Mercury 3 fact | What Inception says |
| --- | --- |
| Release date | October 19, 2026 |
| Model type | Diffusion LLM (dLLM) |
| Speed | 1,068 tokens per second on Nvidia GPUs |
| Context window | 1 million tokens (Mercury 2.5 had 260,000) |
| Price | $0.20 input / $0.70 output per million tokens |
| Launch promo | 50% off: $0.10 input / $0.35 output |
| Free tier | 100 million free API tokens for new accounts |
| Where | Inception API, OpenRouter, Baseten, Vercel, chat.inceptionlabs.ai |

## Mercury 3 price: is it free?

Mercury 3 is a paid API model, but you can try it without paying. Unite.AI reports that new Inception platform accounts get 100 million free tokens with no payment details required. Inception also runs a chat page at chat.inceptionlabs.ai where you can talk to the model in a browser.

After that, the list price is $0.20 per million input tokens and $0.70 per million output tokens. For launch, Inception is taking 50% off, so it is $0.10 and $0.35 for now. Cached input, meaning text the model has already seen in the same session, is 90% cheaper than regular input.

That puts it in the same bargain bin as the other fast models we covered this month. Here is how it lines up, using each company’s published API price:

| Model | Input (per 1M tokens) | Output (per 1M tokens) | Type |
| --- | --- | --- | --- |
| Mercury 3 (list) | $0.20 | $0.70 | Diffusion |
| Mercury 3 (launch promo) | $0.10 | $0.35 | Diffusion |
| MiniMax M3.1 Flash | $0.15 | $0.60 | Autoregressive |
| GPT-5.6 Luna | $1 | $6 | Autoregressive |

One thing to keep in mind: a cheap token is not always a cheap answer. A model that needs more tries or longer replies to get something right can cost more in the end. Real costs depend on your task.

## Mercury 3 benchmarks: how good is it?

This is where you should read carefully. Inception says Mercury 3 matches or beats OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5 on quality, and that its scores on standard tests are three times higher than Mercury 1 mini. But the launch post does not list the individual benchmark scores behind those claims. Until independent testers publish numbers, treat them as company claims.

The more concrete evidence comes from early customers Inception quoted:

- **TinyFish** (browser agents): on 50 live web navigation tasks from the Mind2Web test, Mercury 3 finished 50%, which Inception says was well ahead of GPT-6.1-Sol at a lower per-token cost.

- **Octen AI** (search for smart glasses): median generation time of 0.9 seconds and 2.2 seconds end to end. Its CEO put it simply: "an answer that takes five seconds feels broken."

- **RevisionDojo** (studying): hints and marked answers come back in about a second.

- **Lorikeet** (customer support): led other models on speed in early testing.

Notice the pattern. These are jobs where waiting hurts: voice, search, support chats and agents that click through websites step by step. That is the sweet spot Inception is aiming for.

## Mercury 3 vs Mercury 2

Mercury 3 replaces the older models fast. Mercury 2 and Mercury 2.5 are now closed to new users, and existing customers who used them before October 19 keep access until November 15, 2026. The big upgrades are the larger model, the 1 million token context window (up from 260,000 in Mercury 2.5), and a lower list price than the $0.25 input and $0.75 output that price trackers list for Mercury 2.

Inception also mentioned two siblings in the same post: Mercury Voice, tuned to start speaking answers quickly, and Mercury Decide, a model for high-volume classification. Inception claims Mercury Decide beats [Jev](https://whatisagenticai.net/blog/what-is-jev-ai/) on JevBench and LegalBench, again without listing scores.

## How to try Mercury 3

- **Just chat:** open chat.inceptionlabs.ai in a browser.

- **Build with it:** sign up on the Inception API platform and use the 100 million free tokens. It is OpenAI API compatible, and the model ID is mercury-3.

- **Use a platform you already have:** it is also on OpenRouter, Baseten and Vercel.

## What this means for you

For most people, the interesting part of Mercury 3 is not the word "diffusion." It is what speed does to AI agents. An agent does not answer once. It plans, takes a step, checks, and goes again, sometimes dozens of times. If each step takes five seconds, a task drags. If each step takes one, the same task feels instant. Faster and cheaper models are what make [agentic AI](https://whatisagenticai.net/what-is-agentic-ai/) feel normal in daily life.

You do not need to sign up for an API to benefit. **Agentic AI** puts a team of ready-made expert agents on your iPhone and iPad, running on frontier models, so you skip the model shopping and get the finished email, plan or research summary. You can put up to four agents on one task, and there is no account to create.

## Quick answers

**What is Mercury 3?** Mercury 3 is a diffusion-based large language model from Inception Labs, released October 19, 2026. It writes many tokens at once instead of one by one, reaching about 1,068 tokens per second.

**Who made Mercury 3?** Inception, a startup that builds diffusion language models. It also made Mercury 2, Mercury 2.5 and Mercury Voice.

**Is Mercury 3 free?** New API accounts get 100 million free tokens, and you can chat with it at chat.inceptionlabs.ai. After that, it costs $0.20 input and $0.70 output per million tokens, with 50% off during the launch.

**Is Mercury 3 better than Claude Haiku 5.5?** Inception says it matches or beats Claude Haiku 5.5 and GPT-6 Luna on quality, but it has not published the scores. It is clearly faster. Wait for independent tests before calling it better.

**What is the Mercury 3 context window?** 1 million tokens, up from 260,000 in Mercury 2.5.

**Is Mercury 3 open source?** Inception has not said it is releasing the weights. It is offered through its API and partner platforms.

New to the vocabulary? Our [agentic AI glossary](https://whatisagenticai.net/agentic-ai-glossary/) explains terms like tokens, inference and context window in plain words.

**Sources:** [Inception: Diffusion at Scale, Introducing Mercury 3](https://www.inceptionlabs.ai/blog/introducing-mercury-3) · [Unite.AI](https://www.unite.ai/inception-releases-mercury-3-its-flagship-diffusion-llm/) · [Inception blog (Mercury 2, 2.5 and Voice dates)](https://www.inceptionlabs.ai/blog)