Inception drops Mercury Voice, a 320ms diffusion LLM for voice agents

Anyone who has interacted with an AI phone agent knows the painful silence after finishing a sentence. If an assistant takes longer than 500 milliseconds to jump in, natural human rhythm breaks completely.

Two weeks after teasing Mercury 2.5, Inception Labs officially launched Mercury Voice for enterprise customers today. It is a diffusion-based LLM engineered specifically to kill conversational latency without forcing developers to sacrifice reasoning quality.

Breaking the 500ms voice barrier

Mercury Voice clocks a median time to first answer token of 320 milliseconds on production customer-service prompts, making it 5.9 times faster than GPT-6 Luna. Its 95th-percentile tail latency sits at 750 milliseconds, comfortably ahead of rivals whose tail latency often spikes past 4 seconds.

Built on a diffusion architecture instead of standard autoregressive generation, the model handles tool use, long system prompts, and up to 128K context with 50K output tokens. Builders can choose between three reasoning effort levels: low, medium, and high.

Despite the speed, Inception says the model does not sacrifice intelligence. Across a benchmark suite spanning τ³-bench Telecom, Retail, and Airline, alongside IFBench and BFCL v4, Mercury Voice beat out Gemma 4 31B, GPT-6 Luna, GLM-5.3-Flash, Gemini 3.5 Flash-Lite, and Qwen3.5-397B.

Early enterprise deployments highlight the jump. Audivi AI is running it in drive-thru lanes to handle complex order modifications and upsells without human intervention. Financial voice agent Altur reported that the language model is no longer the bottleneck in their phone stack. OpenCall noted that across their live customer workloads, median response latency dropped to roughly 170 milliseconds.

The model is available as an OpenAI-compatible endpoint, making it a drop-in replacement for frameworks like LiveKit, Pipecat, Vapi, and Retell. Standard pricing sits at $0.40 per million input and $1.50 per million output tokens, but a 50% launch discount brings it to $0.20 input and $0.75 output. That translates to roughly $0.009 per minute of voice conversation, compared to around $0.045 per minute on GPT-4.1.

Why it matters

Voice agent builders have been stuck in an awful trade-off: pick a smart reasoning model and suffer awkward two-second pauses, or pick a fast, dumb model that breaks the moment a caller goes off-script.

Mercury Voice looks like a genuine architecture win. By leveraging diffusion LLMs for text generation, Inception is sidestepping the sequential token generation bottleneck that plagues standard autoregressive transformers. Delivering 320 ms median latency alongside solid tool-calling benchmarks shows that low latency does not have to mean lower IQ.

For enterprise call centers and drive-thru operators, this is definitely worth testing right away. At under a penny per conversation minute, the economics make aggressive agent automation far more attractive.

Still, keep in mind that the LLM is only one slice of the voice loop. Your overall pipeline will still be bounded by speech-to-text and text-to-speech latencies. But by taking the central brain off the critical bottleneck list, Inception just forced rivals like OpenAI and Google to rethink how their lightweight models operate under real-time constraints.