Inception launches Mercury Voice, a diffusion language model aimed at voice agents

Inception launched Mercury Voice for enterprise customers on 29 September 2026. The model is a diffusion-based language model that, per the company, refines multiple tokens in parallel rather than generating them one at a time.

Company-reported figures put median time to first answer token at 320 milliseconds and 95th-percentile latency at 750 milliseconds. Launch pricing is $0.20 per million input tokens and $0.75 per million output tokens, with standard pricing at $0.40 and $1.50 respectively; Inception estimates roughly $0.009 per conversation minute for typical voice-agent workloads. The company says Mercury Voice scored above models including GPT-6 Luna, Gemma 4 31B, GLM-5.3-Flash and Qwen3.5-397B on composite benchmarks, claims more than twice the speed of those comparisons, and says only Mercury Voice achieved a median below 500 milliseconds. Access requires contacting the company. Customer results cited for Audivi AI, Altur and OpenCall are described as reflecting particular deployments rather than guaranteed performance.

Why it matters for digital humans: Time to first token is the budget item that decides whether a talking avatar feels like a conversation or a kiosk, because the face cannot start moving until the language model starts producing. Shaving the text stage toward 300ms leaves room for speech synthesis and frame generation inside the roughly one-second window where turn-taking still reads as natural — which is why the diffusion-LLM approach is worth tracking even for teams with no interest in phone agents. Every latency and benchmark number here is company-reported, and the timing figures measure model response, not the full voice call; treat them as a claim to test in your own pipeline rather than a result.