Hopper Research is an inference and post-training service for companies running real-time voice agents. It takes open-weight speech-to-text, text-to-speech, and speech models, post-trains them on a customer's production calls, then serves them and keeps improving them against live traffic. The founders are launching now because hosted inference is still built around coding agents, leaving voice products on closed models that are slow to start a reply and expensive as call volume grows. Their site lists Gemma 4 31B at 80 ms time to first token, compared with 600 ms for GPT-4.1, and at $0.30 per million input tokens, compared with $2. The Y Combinator launch post also lists Qwen 3 TTS at 150 ms time to first audio, which they describe as 10 times less expensive than ElevenLabs, and Nemotron 3.5 ASR as twice as fast and eight times less expensive than Deepgram Nova-3. Zero data retention is the default. Pavan Katta and Jashwanth Pedapudi are the founders, a two-person team in Y Combinator's Fall 2026 batch in San Francisco. Katta previously ran voice inference at Salient, powering millions of calls a day. Pedapudi wrote inference kernels for Quest 3 at Meta Reality Labs. They studied electrical engineering together at IIT Madras. The company is backed by Y Combinator and by angels including Vijay Krishnan, CTO of Turing, along with people from xAI, Stripe, and Rubrik.