# Inception releases Mercury 3, claiming it beats GPT-6 Luna and Claude Haiku 5.5 on speed

published: 2026-10-21T11:05:30.478Z
topic: infrastructure
source: 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/
url: https://aiagentnews.io/stories/inception-releases-mercury-3-claiming-it-beats-gpt-6-luna-and-claude-haiku-5-5

Inception AI has released Mercury 3, a diffusion language model that it says beats OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 on eight of 12 of its own benchmarks while running four to eight times faster. It lists at 20 cents per million input tokens and 70 cents per million output tokens, and a launch promotion halves both rates for now.

Inception AI Inc. released Mercury 3 as its new flagship model. In the company's benchmarks, it beat GPT-6 Luna and Haiku 5.5 on eight of 12 tests while running four to eight times faster. Inception measured Mercury 3 at 1,068 tokens per second on Nvidia GPUs, compared with about 240 for Haiku 5.5 and about 130 for GPT-6 Luna. Mercury models draft the full response first and refine every position in parallel over a few passes, instead of generating one token at a time.

Mercury 3's widest lead came in coding, where it scored 51.1 on SWE-Bench Atlas against 33.3 for GPT-6 Luna and 32.3 for Haiku 5.5. It trailed Haiku 5.5 on the DeepSWE coding benchmark and on one long-context test. Inception has not disclosed a parameter count, saying only that the model is more than 40 times the size of Mercury 1, and its context window grows to 1 million tokens. No outside group has published results yet.

Developers can reach Mercury 3 on Inception's own platform or through OpenRouter, Baseten and Vercel, and code written for OpenAI's API needs only minor changes to call it. Early users Octen AI and RevisionDojo report median generation times of about a second in their products. Inception was founded in 2024, and Reuters reported in May that Microsoft was in talks to acquire it, though no deal has been announced.

## Why it matters if you build agents

Teams building latency-sensitive features such as voice, smart glasses or tutoring now have another fast model to test against their current choice. Because the API is compatible with OpenAI's format, trying it may take little more than a configuration change. The speed and benchmark claims come from Inception itself, so check coding and long-context results on your own workloads before routing critical tasks to it, and note that the lower launch pricing is temporary.