Mercury 3 Reviewed: A Chatbot That Writes Its Answer in Chunks, Not Word by Word

Inception has launched Mercury 3, which it calls its most capable “diffusion” language model. Most chatbots write one word-piece (a “token”) at a time, each depending on the ones before. Diffusion models instead draft and refine several parts of an answer in parallel, a bit like an editor improving a whole page at once instead of a writer typing left to right. The practical promise is speed: Inception reports 1,068 tokens per second on NVIDIA GPUs, which is several times faster than many standard chatbots.

A review by The Neuron’s Corey put it through eight small tests via OpenRouter for a total cost of about four cents. The results were mixed in a useful way. Business maths was correct apart from a minor mix-up between profit per cup and percentage margin. A short story was funny but overshot its word limit and drifted from the one-room setting. A small dungeon-crawler game passed a syntax check and 2,000 randomised map checks. It rejected a made-up NASA mission, though it added a wrong claim about Shuttle launches. The reviewer calls these informal spot checks, not a proper benchmark.

The specs are attractive. The context window (how much text it can keep in view at once) is 1 million tokens, up from 260,000 in the previous version. Standard prices are $0.20 per million input tokens and $0.70 per million output tokens, with a launch discount of half that. Inception says it matches or beats GPT-6 Luna and Claude Haiku 5.5 on several tests, for example 55.18 versus 50.50 and 45.78 on BrowseComp+, though Haiku leads on some and Luna on one long-context test. These are company-reported numbers. New API accounts get 100 million free tokens.

What’s behind it: speed matters more than it sounds. When an AI agent has to do dozens of small steps, such as browsing, checking and retrying, each slow step adds up, and cheap fast models make those loops practical. The trade-off with diffusion so far has been precision. The review notes imperfect handling of constraints and some overconfident explanations, which is the sort of thing you only trust after independent testing.

What this means for you: if you use AI casually, you will not notice any of this today. If you build apps or automate tasks, the free token allowance makes Mercury 3 an easy, nearly costless experiment, especially for jobs where speed beats perfection, such as drafting, classifying or search helpers. Keep a stronger model for anything where a small factual slip would hurt.

Sources

• The Neuron: Mercury 3 review, diffusion language models get serious
• The Neuron: Everything that happened in AI today, October 20