Hopper launches voice-agent inference that trains on your own call recordings

A voice agent is only as good as its response time. The reader of this item has probably sat through a two-second pause while a model thinks of its first word. The pause ends a conversation. Hopper wants to end the pause.

Hopper launched out of Y Combinator Fall 2026 on 26 October. The YC launch page received 24 votes in its first hours. The company offers inference for voice agents. It also offers to train open-weight models on a company's own production call recordings. The pitch is lower latency and lower cost than the main third-party providers, Hopper says.

The latency numbers Hopper claims

Hopper lists three models and three latency figures on its launch. The large language model is Gemma 4 31B, with 80 ms time-to-first-token. The text-to-speech model is Qwen 3 TTS, with 150 ms time-to-first-audio. The speech-to-text model is Nemotron 3.5 ASR. Every number is Hopper's own measurement.

The company compares those figures against named vendors. For the language model, Hopper says 7x faster and 4x cheaper than GPT 4.1. For speech, it says 10x cheaper than ElevenLabs. For transcription, it says 2x faster and 8x cheaper than Deepgram Nova-3. Every comparison is a vendor measuring itself against competitors. No third party has run these numbers.

Why training on production traces is the real pitch

The headline numbers are latency and cost. The founders say the deeper gain is in training. A base voice model has never heard a company's product names, its customer names or its internal terms. Hopper trains the open-weight model on the company's actual call recordings. It then serves and improves the model against live traffic, the launch says.

The named benefits on the page are better instruction-following, better tool-calling and better pronunciation on domain-specific words. A base model may say a customer's name wrong every third sentence. That is a product problem on its own, and no amount of extra speed solves it.

Who built it

Hopper's team includes Pavan Katta. He ran voice inference at Salient and powered millions of calls a day there, the launch page says. Jashwanth wrote inference kernels for Quest 3 at Meta Reality Labs. The company is based in San Francisco.

What is missing from the launch

The launch page lists no customer logos and no third-party benchmark. The claim that Hopper is 7x faster than GPT 4.1 is Hopper's own. The claim that it is 2x faster than Deepgram Nova-3 is Hopper's own. A voice team that lives on sub-second latency has to re-run the test on its own calls. Its own hardware, its own prompts, its own numbers.

The product is called "fast, open-source inference for real-time agents" on the YC listing. Whether the service, the models or both are open source is not stated on the launch page. The company's own site is at withhopper.com.

Source

• Primary source: Hopper on Y Combinator.
• Product site: Hopper.