Phonic-1 Tops Speech-to-Speech Rankings, Calls Tools While You Talk

Anyone who has called a support line knows the phrase: "Sure, let me look that up for you." Then comes the silence. Voice AI startup Phonic wants to delete that silence entirely.

On October 29, Phonic launched its new model, Phonic-1, alongside a $27.5 million Series A led by First Round, with Lux Capital, Felicis, Definition and Essence VC participating. Total funding now stands at $31.5 million, and the angel list is stacked: Clement Delangue of Hugging Face, Guillermo Rauch of Vercel, Amjad Masad of Replit and Naval Ravikant among them.

The answer is ready before the caller stops talking

Phonic-1's headline trick is calling tools while the caller is still speaking. Conventional voice agents wait for the person to finish, say "let me check," and only then hit an API. Phonic-1 works as it listens, so the result is in hand the moment the caller goes quiet.

It sounds like a minor optimization. On a real phone call it is the difference between a good and a bad experience. Checking an order, moving an appointment, verifying a policy: the bottleneck is rarely how fast the model thinks, it is the serial chain of listen, think, call, answer. Pulling the tool call into the listening phase hides the wait inside the conversation itself.

On the leaderboard, Phonic-1 took first place on the Artificial Analysis Speech to Speech Index at 87.1, a hair above Grok Voice 3.0 at 86.9. A narrow lead, but enough for a headline. Time to first audio is about 0.61 seconds, in line with the company's 600 ms claim. On Sierra's multilingual full-duplex benchmark, τ³-Voice, it scored 83.2, ahead of GPT-Live-1 at 79.9 and Gemini 3.8 Live at 68.5.

End to end instead of stacked parts

Most voice agents today are a three-piece cascade: one model transcribes, one reasons, one speaks. Cascades are easy to build and tune, but every transcription throws information away: pauses, hesitation, irritation, and exactly how someone pronounced their name.

Phonic has done nothing but speech-to-speech for more than two years, rebuilding data, pre-training, RL, kernels and infrastructure along the way. Phonic-1 hears raw audio, picks up tone and hesitation, handles interruptions and turn-taking, and holds up on noisy phone lines. It covers 50+ languages, can switch languages mid-call in the same voice, and has been tuned to pronounce URLs, emails and alphanumeric strings correctly, precisely where support calls tend to go wrong.

Phonic doesn't pretend a voice model can carry all the reasoning alone, either. Frontier models act as supervisor agents that help with tool calls and harder reasoning. It is a pragmatic split: the end-to-end model sounds human, the big model keeps it from being wrong.

Notably, Phonic admits it still trails real people on naturalness, with humans winning roughly 60% of blind comparisons. That kind of candor is rare in a launch post, and it suggests the company knows the next round of competition won't be about latency alone.

The business is already running

Pricing is a single per-minute rate starting at $0.07, with no LLM pass-throughs or per-token fees. For enterprise buyers that is far simpler than a cascade where three vendors each send their own bill. Compliance boxes are ticked: SOC 2 Type II, HIPAA, zero data retention, a 99.9% uptime SLA, and self-hosted deployment.

Phonic says it already handles calls on six continents for customers including Ashby, Maven AGI, Rho and a Fortune 50 IT infrastructure company, across support, healthcare, recruiting and logistics.

Voice agents are shifting from "can it talk" to "can it get the job done." OpenAI, Google and xAI are all pouring resources into voice, and a startup can't outmuscle them on general capability. But on the specific problem of finishing a task over the phone, acting while listening is a gap the giants haven't closed yet. How long that window stays open depends on how fast they follow.