Inception Debuts Mercury 3: Diffusion-Based AI Outperforms GPT-6 Luna at 1,000+ Tokens Per Second

AI startup Inception has announced the launch of Mercury 3, a new language model based on diffusion architecture rather than the standard auto-regressive method used by industry leaders. According to the company, Mercury 3 outperforms top-tier models like GPT-6 Luna and Claude Haiku 5.5 in speed and accuracy across several key benchmarks, operating at four to eight times the speed of its competitors.

A Shift in Architecture: From Words to Drafts

Unlike traditional models such as GPT or Claude, which generate text word-by-word (auto-regressive), Mercury 3 utilizes a diffusion process. This method, commonly associated with image generation models, involves creating a rough draft of the entire response and refining it through several simultaneous passes.

This architectural shift allows Mercury 3 to generate more than 1,000 tokens per second on Nvidia GPUs. In internal comparisons, Inception reported that Mercury 3 reached 1,068 tokens per second, significantly outpacing Claude Haiku 5.5 (approximately 240 tokens per second) and GPT-6 Luna (approximately 130 tokens per second).

Benchmarks and Performance

Inception’s internal testing shows Mercury 3 leading its rivals in eight of 12 tests. The model demonstrated a significant advantage in instruction following and autonomous coding:

• IFBench (Instruction Following): Mercury 3 scored 76.84 points, surpassing GPT-6 Luna’s 61.84.
• SWE-Bench Atlas (Coding): Mercury 3 achieved 51.08 points, while its rivals scored between 32.26 and 33.33.
• Web Navigation: At TinyFish AI, Mercury 3 solved 25 out of 50 navigation tasks on real websites, reportedly beating GPT-6.1-Sol.

However, the results were not a total sweep. Claude Haiku 5.5 maintained a lead in the DeepSWE long coding test and information retrieval from long texts. Additionally, GPT-6 Luna held a slight edge in two tests conducted by Artificial Analysis.

Real-Time Applications and AI Agents

Inception is positioning Mercury 3 as the primary solution for applications where latency is a critical bottleneck, such as autonomous AI agents, smart glasses, and real-time customer support.

Early-access partners have highlighted the impact of this speed. Kuan Zou, founder of smart glasses firm Octen AI, noted that previous language models were a bottleneck for wearable tech, where a five-second delay feels "broken." With Mercury 3, the median response time dropped to 0.9 seconds. Similarly, education company General Learning reported delivering hints and corrections to students in roughly one second.

Pricing and Availability

Mercury 3 features a 1-million-token context window and uses an OpenAI-compatible API, allowing developers to switch models with minimal code changes.

The standard pricing is set at $0.20 per million input tokens and $0.70 per million output tokens. At launch, Inception is offering a 50% discount, though the company has not specified when this promotional period will end. While the post-discount output price is higher than the $0.50 charged by GPT-6 Luna and Claude Haiku 5.5 for short inputs, Mercury 3 remains more cost-effective for very long prompts.

The model is currently available through Inception’s platform, as well as OpenRouter, Baseten, and Vercel. Inception, which previously raised $50 million, counts Nvidia’s NVentures, Microsoft’s M12, and Snowflake Ventures among its backers.