Inception Labs ships Mercury 2.5, a diffusion LLM pitched at latency-sensitive production workloads
1,107 tokens per second on standard Nvidia GPUs, a 260K context window, and an 80%-off launch price of $0.04/$0.15 per million tokens, aimed at voice agents, coding subagents, and support workloads where speed and cost matter more than leaderboard rank.
Inception Labs launched Mercury 2.5 on September 9, 2026, the latest in its Mercury family of diffusion-based language models. Diffusion models generate text by iteratively refining a full draft instead of producing it token by token, which is where the speed claims come from. The company says Mercury 2.5 delivers a 40% intelligence gain over its predecessor, runs at 1,107 tokens per second on standard Nvidia GPUs, and supports a 260K-token context window, with intelligence roughly comparable to GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5 on the benchmarks Inception cites. Standard pricing is $0.20/$0.75 per million input/output tokens; launch pricing cuts that 80%, to $0.04/$0.15.
Speed is the pitch
CEO Stefano Ermon called it "our most capable production model yet," but the company's positioning leans hardest on latency. OpenCall's Oliver Silverstein, quoted in the launch post, said switching to Mercury meant "our P99 response time dropped from several minutes to just one second." That is a named customer with a specific number.
The unit-economics argument
Diffusion LLMs have been a small lane next to the dominant autoregressive models. Mercury 2.5 targets cost-conscious, latency-sensitive production traffic: voice agents, coding subagents, customer support. If you pay per token for a slower frontier model on a workload where speed is the product requirement, benchmark it against a sub-second P99 claim at the 80%-off launch price.