Inception Labs Launches Mercury 3: A Diffusion LLM Challenging GPT-6 Luna and Claude Haiku 5.5

Inception Labs has released Mercury 3, a new diffusion language model that the company claims surpasses OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 in 8 of 12 benchmark tests, delivering significantly faster generation speeds and lower costs.

Mercury 3's Diffusion Architecture and Performance Claims

Mercury 3 distinguishes itself through its diffusion architecture, which generates text by refining a rough draft across multiple passes, rather than the sequential token generation used by models like GPT and Claude. This approach reportedly contributes to its speed and cost advantages. Inception Labs states that Mercury 3 achieves generation speeds of 1,068 tokens per second on Nvidia GPUs, positioning it as roughly 4 to 8 times faster than GPT-6 Luna and Claude Haiku 5.5.

The company's internal benchmark suite indicates that Mercury 3 wins 8 out of 12 tests against its competitors. Specifically, Mercury 3 achieved an instruction following (IFBench) score of 76.84, approximately 15 points higher than GPT-6 Luna. Its autonomous coding (SWE-Bench Atlas) score is reported at 51.08. While Mercury 3 leads in several areas, Claude Haiku 5.5 reportedly excels in the long DeepSWE coding test, and GPT-6 Luna leads in two Artificial Analysis tests. these benchmarks are vendor-run and have not yet received independent verification.

Availability and Pricing Structure

Mercury 3 is currently available on Inception Labs' own platform, as well as through OpenRouter, Baseten, and Vercel. The launch pricing for Mercury 3 is set at $0.20 per million input tokens and $0.70 per million output tokens, which represents a 50% reduction from its regular pricing. This pricing model aims to make the model more accessible, especially for applications requiring high throughput.

Inception Labs highlights that Mercury 3 has more than 40 times the parameters of its predecessor, Mercury 1, while maintaining the same operational speed. This suggests a significant scaling capability for diffusion LLMs, similar to what has been observed with autoregressive models.

Implications for Agentic AI and Real-World Applications

The primary target for Mercury 3 is agentic AI, where numerous model calls are chained together, making even small delays impactful. Early partners have reported practical applications demonstrating Mercury 3's low latency. For instance, smart-glasses search answers were delivered with a median response time of 0.9 seconds, and tutoring hints were provided in approximately one second. These examples suggest that Mercury 3's speed could be beneficial for real-time AI applications.

The emergence of diffusion LLMs like Mercury 3, alongside Google's Gemini Diffusion and DiffusionGemma, indicates a growing architectural challenge to the established autoregressive models. This development could lead to new possibilities for AI performance and efficiency across various domains.

Conclusion

Inception Labs' Mercury 3 represents a notable entry into the competitive LLM landscape, leveraging a diffusion architecture to offer claimed advantages in speed and cost. While the company's internal benchmarks suggest strong performance against leading models like GPT-6 Luna and Claude Haiku 5.5, independent verification will be crucial for broader industry acceptance. Its availability across multiple platforms and competitive launch pricing make it a model to watch for developers and organizations focused on agentic AI and real-time applications.

Sources

• arXiv:2506.17298v1 [cs.CL] 17 Jun 2025
• Add Inception Labs (Mercury) as a model provider · Issue #6458 · zeroclaw-labs/zeroclaw · GitHub
• Beyond Autoregression: An Empirical Study of Diffusion Large Language Models for Code Generation
• Mercury: Ultra-Fast Language Models Based on Diffusion
• GitHub - hamzaamjad/mercury-client: Python SDK for Inception Labs Mercury diffusion-LLM API · GitHub