Mercury Voice

Mercury Voice is a diffusion language model for real-time voice agents that Inception announced on its blog on 29 September 2026. Inception’s Mercury models generate text by diffusion, refining many tokens in parallel, rather than one token at a time. According to the announcement, Mercury Voice takes up to 128,000 input tokens and produces up to 50,000 output tokens, is built for voice-agent turns, and is available to enterprise customers through Inception’s OpenAI-compatible API, by contact with its sales team. Inception gives a median time to first token of 320 milliseconds and a 95th percentile of 750 milliseconds, and compares its speed favourably with a leading mid-tier frontier model; these are the vendor’s own measurements. The list price is $0.40 per million input tokens and $1.50 per million output tokens, with a 50% launch discount bringing that to $0.20 and $0.75, which Inception puts at about $0.009 per minute of conversation. Inception also cites results on a conversational-agent benchmark without publishing figures that could be checked. The board already lists Inception’s Mercury 2.5.

Not yet independently verified: Every detail here, including the release date, the speed figures and the price, comes from Inception’s own announcement; no independent report or measurement of Mercury Voice was found. We will update this when it can be confirmed, and remove this note.