Mercury 2: The First Reasoning Diffusion Language Model — 5x Faster Than Claude Haiku at Half the Cost

TamilTech AI summary

Inception Labs launched Mercury 2 on February 24, 2026, calling it the world’s first Reasoning Diffusion Language Model, and it generates text in a totally different way from ChatGPT or Claude. Instead of writing one token at a time like normal transformers, it starts with a rough noisy draft of the whole answer and refines many tokens in parallel through denoising steps, which is why it hits about 1,009 tokens per second—roughly 5x faster than Claude 4.5 Haiku—while costing far less per million tokens. That speed and price cut matter a lot for real-time work like agentic coding, customer support at scale, multi-step AI agents, and high-volume India use cases such as vernacular assistants, EdTech tutoring, fintech checks, and government services. You should know the company says coherence issues common to diffusion language models are largely fixed, though independent tests are still ongoing and long-form creative writing may still favor classic transformers for now. API access is live at inceptionlabs.ai, so if you’re weighing AI infrastructure costs and latency, Mercury 2 is worth a serious look.

• What is Mercury 2 and how is it different from ChatGPT or Claude?
• How fast is Mercury 2 compared to other AI models?
• Can Indian developers access Mercury 2's API?
• What are the limitations of Mercury 2's diffusion approach?

AI-assisted summary, checked by the TamilTech editorial team.