Hopper

Companies building voice AI products need speech-to-text, text-to-speech, and speech LLM models that respond quickly enough for natural conversation, but general-purpose models like GPT-4.1 are too slow and expensive for this use case. The page cites 600 ms time-to-first-token and $2 per 1M input tokens as the comparison point that voice applications are stuck with. This latency and cost makes it harder to build responsive, production-grade voice agents.

Hopper trains STT, TTS, and speech LLM models on a customer's own production call data, then serves and continuously improves those models against live traffic. The company highlights specific performance numbers: 80 ms time-to-first-token versus 600 ms for GPT-4.1, and $0.30 per 1M input tokens versus $2 for GPT-4.1. It also references a Gemma 4 31B LLM alongside TTS, STT, and audio components, and shares technical writing on topics like megakernels, GPU throughput, and optimizing diffusion TTS and LLM clients for voice agents, suggesting the product centers on inference-speed engineering.