Most AI language models generate text one token at a time, left to right, never revising. Mercury 3 doesn’t work that way. Released by Inception Labs this week, it uses a fundamentally different architecture, and the benchmark results are hard to dismiss.
The dominant design for language models is called autoregressive generation. The model predicts each token (a small chunk of text, roughly a word or part of a word) based on everything before it, then moves forward one step at a time. Mercury 3 is built on diffusion, a technique borrowed from image generation tools. Diffusion models don’t write sequentially. They start with noise and refine the entire output at once. Inception Labs spent years applying this approach to language at commercial scale.
The speed numbers are striking. Mercury 3 generates more than 1,000 tokens per second on NVIDIA GPUs, 4 to 8 times faster than GPT-6 Luna and Claude Haiku 5.5. It doesn’t just win on speed, either. Across standard benchmarks, Mercury 3 scores on par with or better than both models. Speed and quality, not a tradeoff between them.
That combination matters in real applications. Voice assistants become uncomfortable to use after about 400 milliseconds of delay. Browser-use agents, AI systems that navigate websites and complete tasks on your behalf, make dozens of model calls per task, and latency compounds at every step. One production partner tested Mercury 3 on 50 live web navigation tasks. It completed 50% of them. That beat GPT-6.1 Sol, using the same number of steps and tokens, at a fraction of the per-token cost.
The pricing makes the math work. Mercury 3 launches at $0.10 per million input tokens and $0.35 per million output tokens. It also ships with a 1 million token context window. Context window refers to how much text the model holds in working memory at once. One million tokens is enough to process an entire large codebase, a full document library, or an extended conversation history without breaking it into pieces.
The larger story is about architecture. The AI community spent years questioning whether diffusion could ever scale to compete with transformer-based models like GPT and Claude. Mercury 3 is 40 times larger than Mercury 1, three times more capable on benchmarks, and still just as fast. The scaling laws that drove progress in transformers held here too. The experiment worked.
Mercury 3 is available now through the Inception API, OpenRouter, Baseten, and Vercel. It is fully compatible with the OpenAI API, so it drops into existing stacks without a rewrite.
Want to explore how next-generation AI models could benefit your business? Let’s talk.