Inception's Mercury Voice answers in 320ms

Inception released Mercury Voice, a diffusion language model tuned for voice agents, and made it generally available to enterprise customers. A diffusion LLM generates several tokens at once rather than one at a time, which is where the latency claim comes from.

The measured numbers are 320 ms median time to first answer token on a test set of real customer service prompts, with p95 at 750 ms. Inception calls it 5.9 times faster than GPT-6 Luna without reasoning, and states it is the only model tested with a median under the 500 ms budget that natural conversation implies. It also says its p95 is faster than the median of every model it tested except GPT-OSS-120B at low effort on Cerebras.

Specs are 128K context with up to 50K output tokens and three reasoning effort settings, low, medium and high. Pricing is $0.40 per million input and $1.50 per million output, at 50% off for launch, which works out to $0.20 and $0.75. Inception’s worked example is about $0.009 per minute of conversation, roughly five times cheaper than GPT-4.1 at $0.045 per minute.

On quality, Inception reports beating GPT-6 Luna, GLM-5.3-Flash, Gemini 3.5 Flash-Lite, Qwen3.5-397B and Gemma 4 31B on a composite of the tau cubed bench Telecom, Retail and Airline suites plus IFBench and BFCL v4, and says it is about twice as fast as the next fastest model on the conversational set. The comparisons are Inception’s own against models they tested, so treat them as directional.

It ships as an OpenAI API compatible endpoint, which means it drops into the LLM slot of LiveKit, Pipecat, Vapi, Retell or an in-house stack. Named customers include Audivi AI in drive-thru ordering, Altur for financial voice agents, and OpenCall, which cites roughly 170 ms median model latency on its production workload.

Two things to weigh. The model is enterprise only, requiring a conversation with Inception’s sales team, so this is not something to try tonight. And on tail latency, Inception’s own framing is that a p95 of four seconds means roughly one turn in twenty fails, which puts 750 ms in context as good but not trivial. The page also carries two conflicting dates, a frontmatter timestamp of October 1 and a blog listing of September 29.

(Original source: Inception)