# Mercury 3 writes the whole answer at once, and fast

URL: https://ai-pulse.aglaganova.com/blog/mercury-3-writes-the-whole-answer-at-once-and-fast
Published: 2026-10-22
Author: Roman Aglaganov, Full-stack developer
Category: Models
Region: North America
Tags: benchmarks, coding

> Inception’s new diffusion model drafts an entire response and refines it in parallel, rather than writing one word at a time. The company says it beats GPT-6 Luna and Claude Haiku 5.5 on most of its tests at several times the speed.

![Illustration: a row of blurred grey blocks sharpening from left to right into crisp black blocks, with one red block](https://ai-pulse.aglaganova.com/api/media/file/mercury-4.png)

Almost every chatbot writes its answer one token, roughly a word fragment, at a time. Inception, a start-up founded in 2024 by the Stanford professor Stefano Ermon and fellow academics, builds “diffusion” language models that work differently: they draft the whole response up front, then refine every part of it in parallel over a few passes. Image generators work in a similar way. On 19 October it released Mercury 3, its biggest model so far.

## What Inception claims

**Speed:** 1,068 tokens a second on Nvidia graphics chips, against about 240 for Claude Haiku 5.5 and 130 for GPT-6 Luna in the same comparison.

**Quality:** better than both rivals on eight of 12 benchmarks. Its biggest lead was in coding, scoring 51.1 on SWE-Bench Atlas against 33.3 for GPT-6 Luna. It trailed Haiku 5.5 on DeepSWE, a second coding test, and on one long-context test.

**Size:** Inception has not given a parameter count, saying only that the model is more than 40 times the size of Mercury 1. It can read up to 1 million tokens, up from 260,000 for Mercury 2.5.

All these figures come from Inception, and no outside group has published its own results yet.

## Price and access

Mercury 3 lists at $0.20 per million input tokens and $0.70 per million output tokens, half that during a launch offer, which is more than the list price of the two rivals it compares itself with. It is available through Inception’s own service and through OpenRouter, Baseten and Vercel, and works with software written for OpenAI’s API. The weights are not released.

## Why it matters

The main doubt about diffusion has been whether it scales. “Compute and data, not fundamental algorithmic barriers” are the limits, Ermon told SiliconANGLE. If he is right, the speed matters most where people notice every delay: early customers are using it for voice assistants, smart glasses and tutoring apps.

## Sources

- [SiliconANGLE: Inception says Mercury 3 beats GPT-6 Luna and Claude Haiku 5.5 at up to 8x the speed](https://siliconangle.com/2026/10/20/inception-says-mercury-3-beats-gpt-6-luna-and-claude-haiku-5-5-at-up-to-8x-the-speed/)
- [The Decoder: Inception’s Mercury 3 diffusion model claims to beat GPT-6 Luna and Claude Haiku 5.5](https://the-decoder.com/inceptions-mercury-3-diffusion-model-claims-to-beat-gpt-6-luna-and-claude-haiku-5-5-at-a-fraction-of-the-wait/)
- [Stefano Ermon on X: “Today, we’re releasing Mercury 3”](https://x.com/StefanoErmon/status/2112222911218291051)