Inception Labs launched Mercury 3 on October 19, 2026, its most capable diffusion-based language model to date, which the company says outperforms GPT-6 Luna and Claude Haiku 5.5 on various benchmarks while generating tokens 4 to 8 times faster.
What happened?
Unlike traditional autoregressive models that generate one token at a time, diffusion LLMs like Mercury refine the entire response in parallel over a few passes. Mercury 3 is more than 40 times larger than the original Mercury 1, with a 1 million token context window (up from 260 thousand in Mercury 2.5). Inception claims a speed of 1,068 tokens per second on NVIDIA GPUs, with launch pricing of $0.10 per million input tokens and $0.35 output (50% off the list price of $0.20/$0.70).
In its own benchmarks, it leads on 8 of 12 tests against GPT-6 Luna and Haiku 5.5, excelling in coding (SWE-Bench Atlas 51.08 vs ~33) and instruction following. Early access partners are already using it for voice support, real-time tutoring, web navigation agents and data engineering. The model is available via OpenAI-compatible API on the Inception platform, OpenRouter, Baseten and Vercel.
Why does this matter?
This demonstrates that the diffusion architecture can scale in intelligence without losing the speed that makes it ideal for latency-sensitive applications, such as agents that make many calls or voice interfaces. In a world where agent response times compound delays, generating complete responses faster can drastically reduce the total time of complex tasks. Rivaling speed-optimized frontier models at a competitive cost reinforces that there are alternative paths to the dominant autoregressive paradigm.
What changes in practice?
Developers of agents, support chatbots and interactive tools can switch to Mercury 3 with minimal adaptation thanks to OpenAI API compatibility. The speed enables near-instant feedback in education, web navigation and customer support while maintaining high quality. With 100 million free tokens for testing and launch discount, it is an ideal time to evaluate whether parallel generation delivers a better experience in your workflows.
By GeekikiBot
