Inception AI

Mercury 2.5 is Inception's newest diffusion large language model, aimed at teams building search agents, voice agents, and coding tools where response latency is the bottleneck. Unlike most production LLMs, which generate text autoregressively one token at a time, diffusion LLMs start with a rough draft of the output and refine tokens in parallel. That distinction is the whole pitch here: Mercury 2.5 runs at over 1,100 tokens per second on widely-available NVIDIA GPUs, not specialized silicon , while closing the intelligence gap with the frontier tier. Compared to Mercury 2, the new model jumps 10 points on intelligence benchmarks and lands in the range of cost-optimized frontier models like GPT-5.6 Luna, Gemini 3.5 Flash-Lite, and Claude Haiku 4.5. Pricing settles at $0.20 and $0.75 per million input and output tokens, with an 80% launch discount to $0.04 and $0.15, and the context window extends to 260K tokens with tunable reasoning, native tool use, JSON mode, and parallel tool calls added. Inception is also previewing Mercury Voice, tuned for voice agents with sub-170ms time-to-first-token, and Mercury Router, which dispatches prompts across open and closed models based on quality, speed, and cost. It is available through Inception's API, OpenRouter, and Baseten. Inception is based in Redwood City and was founded by the researchers who did much of the underlying work on diffusion modeling: Stanford professor Stefano Ermon, who co-invented diffusion models, along with UCLA professor Aditya Grover and Cornell professor Volodymyr Kuleshov. The launch matters because it is a real test of whether diffusion can be more than a speed trick, showing an architecture that gets cheaper and smarter in the same release rather than trading one for the other.