Hopper

Hopper provides inference infrastructure for voice AI. The company trains speech to text, text to speech and speech language models on a customer's own production calls, then serves and continuously improves them against live traffic, with a focus on low latency. The vendor's headline figures are an 80 millisecond time to first token on Gemma 4 31B, compared with 600 milliseconds for GPT 4.1, and a price of $0.40 per million input tokens, compared with $2 for GPT 4.1. Its engineering blog covers topics such as megakernels, optimizing time to first token for a hybrid mixture of experts model, tuning LLM clients for voice agents and faster diffusion text to speech. Onboarding can be started by talking to an engineer or by copying an onboarding prompt. It is aimed at companies running voice agents at scale, such as call automation and conversational support, where response speed and model quality on their own call data matter more than generic hosted models. The founders previously worked on voice inference at Salient and on inference kernels at Meta Reality Labs, and the company is backed by Y Combinator. The only published price is the $0.40 per million input tokens figure shown for the Gemma 4 31B language model. No pricing page exists for speech models or other plans, so further costs are not stated.