Inception Mercury 3 diffusion LLM claims 1,068 tokens/sec

Inception launched Mercury 3 on October 19, 2026, calling it the most capable diffusion language model released so far. The company says it runs at 1,068 tokens a second on Nvidia GPUs, reads up to 1 million tokens, and costs $0.20 per million input tokens. Every speed and quality figure comes from Inception or its early customers, so the real test is your own workload.

How a diffusion model writes differently

Most chat models, including GPT and Claude, write one token at a time. A token is a word or a piece of a word, and each one has to wait for the one before it. Inception pitches diffusion as an alternative to that token-by-token approach.

Diffusion is the technique behind many image generators. Instead of writing strictly left to right, the model refines a rough version of the output over several passes. Inception's bet is that this makes text models much faster, and Mercury 3 is its argument that the approach also gets smarter as it grows.

"Mercury 3 demonstrates that text diffusion scales," the Inception announcement says. The company claims the scaling laws that drove ordinary models also hold for diffusion. Scaling laws are the observed pattern that bigger models trained on more data get steadily better.

Stefano Ermon, Inception's CEO and co-founder, went further on X. He called Mercury 3 "by far our most powerful model and the most intelligent diffusion LLM ever released," as quoted by RuntimeWire.

Mercury 3 specs and pricing

Item	Mercury 3
Speed	1,068 tokens per second on Nvidia GPUs (Inception's figure)
Context window	1 million tokens (Mercury 2.5: 260,000)
Size	More than 40 times the total parameters of Mercury 1
List price	$0.20 input / $0.70 output per million tokens
Launch price	$0.10 input / $0.35 output, 50% off, no end date given
Cached input	$0.02 list, $0.01 during the launch offer
Free usage	100 million tokens for new accounts, no card needed
Where to get it	Inception API, OpenRouter, Baseten, Vercel

The API follows OpenAI's format, so existing client libraries work with a new base URL. Unite.AI lists the model ID as mercury-3. Inception says it runs 4 to 8 times faster than GPT-6 Luna and Claude Haiku 5.5.

What the benchmarks show, and what they leave out

Inception says Mercury 3 matches or beats GPT-6 Luna and Claude Haiku 5.5 "across a range of benchmarks." It publishes no named scores for that comparison, and neither outlet that covered the launch adds any. It also says the model is three times smarter than the smallest Mercury 1.

The one concrete agent test came from TinyFish AI, an early customer. It ran Mercury 3 on 50 live Mind2Web tasks, a test where an agent has to complete real jobs on websites. Mercury 3 finished 50% of them. Inception says that is "significantly more" than GPT-6.1 Sol, at similar token use and lower cost per token. It gives no Sol score to compare.

Speed claims from customers are more specific. Octen AI tested the model inside its Sense API. It measured a median of 0.9 seconds to generate an answer and 2.2 seconds end to end. Speaking about answers on smart glasses, Octen founder Kuan Zou put the stakes plainly: "An answer that takes five seconds feels broken."

Mercury 2 and 2.5 close to new users

According to Unite.AI, Mercury 2 and Mercury 2.5 no longer accept new users. Existing customers who used them before October 19 keep access until November 15, 2026.

What this means for developers

Run your own test before you switch. The speed claim is the headline, but the quality comparison has no published numbers. Take 50 to 100 real prompts from your app, run them through Mercury 3 and your current model, and score the answers yourself. The OpenAI-compatible API makes that a short job.

Look at latency-bound features first. Voice assistants, search agents and live tutoring all appear in Inception's customer list, and they share one trait: a slow reply ruins the feature. If your product waits on a model while a user stares at a spinner, that is where a faster model pays off.

Budget at list price, not launch price. The 50% discount has no stated end date, so plan on $0.20 and $0.70 per million tokens. For comparison, MiniMax priced M3.1 Flash at $0.15 and $0.60 this week.

Move off Mercury 2 and 2.5 before November 15. Teams on the older models should test Mercury 3 now rather than in the last week. A 1 million token context is a big change from 260,000, so check that long prompts still behave the way you expect.