Mercury Voice

A voice agent has a fraction of a second to answer before a pause feels awkward, and Mercury Voice is built for that. Inception announced the diffusion language model on its blog on 29 September 2026, for real-time voice agents.

Diffusion instead of word-by-word generation

Inception's Mercury models generate text by diffusion: they refine many tokens in parallel rather than producing one token at a time. That is the basis for the speed Inception claims, and Mercury Voice applies that to the brief exchanges typical of spoken dialogue. The company already lists an earlier model in the family, Mercury 2.5.

Specifications and access

The announcement gives limits of 128,000 tokens in and 50,000 tokens out. It is available to enterprise customers through Inception's OpenAI-compatible API, by contacting its sales team, rather than through open self-serve sign-up.

Speed figures

Inception quotes 320 milliseconds as the median wait for a first token and 750 milliseconds at the 95th percentile. It compares its speed favorably with a leading mid-tier frontier model. It also refers to a conversational-agent benchmark but publishes no numbers that can be verified. These are the vendor's own measurements, and no independent report or measurement of Mercury Voice was found.

Pricing in outline

Inception quotes per-million-token rates for input and output, with a launch discount of 50 percent, and translates that into an approximate cost per minute of conversation. Buyers should confirm current rates with the vendor.