Phonic-1 Starts Calling Tools Before the Caller Stops Talking, and Phonic Raised $27.5M to Sell It

Most voice agents have a tell: the caller asks a question, the agent says “let me look that up,” and only then goes to fetch the answer. Phonic says its new model removes that pause by acting while the person is still speaking. On October 29, 2026, the company introduced Phonic-1, a speech-to-speech model built for agentic phone calls, and announced a $27.5 million Series A.

First Round led the round, with Lux Capital, Felicis, Definition and Essence VC taking part. Phonic says the raise brings its total funding to $31.5 million. Its launch post also lists angel investors including Erik Bernhardsson of Modal, Clement Delangue of Hugging Face, Arash Ferdowsi of Dropbox, Amjad Masad of Replit, Guillermo Rauch of Vercel, Gokul Rajaram, Naval Ravikant and Qasar Younis.

Why Phonic skips the transcript

Phonic’s argument starts with how most voice agents are built today. One model transcribes the caller, another decides what to say, and a third speaks the reply. The company calls that cascade a necessary first step but says “a transcript is a lossy copy of a conversation,” because it drops frustration in a voice, a long pause before an answer, or the way someone pronounces their own name.

Speech-to-speech models keep the audio end to end. Phonic’s claim is that, until now, they forced builders to trade intelligence for naturalness, producing impressive demos that rarely reached production. The team says it has worked only on speech-to-speech since it was founded more than two years ago, and that Phonic-1 came from rebuilding “every layer of the stack: data, pre-training, RL, kernels, and infrastructure.”

Acting mid-sentence

The headline feature is timing. Phonic says Phonic-1 is “the only model that can act while the caller is still talking, so the answer is ready when they stop.” Other models, it says, wait for the caller to finish before calling a tool. The company’s site adds that Phonic uses frontier models as supervisor agents to help with tool calling and reasoning, and quotes a 600 ms response time, against “a full second or more” for other agents.

To show long-call stamina, Phonic published a 44-minute recorded call in which the model, acting for a hotel in Brazil, helps a maid of honour book a bachelorette party. The company says that call involved 123 tool calls, and it shows the transcript, per-turn latency and every tool call in its platform.

Leaderboard claims and a candid human comparison

Phonic says Phonic-1 took the top spot on the Artificial Analysis Speech to Speech Index and sits alone on the Pareto frontier of intelligence against latency in Artificial Analysis’s full-duplex tests. It cites charts sourced to Artificial Analysis, Sierra’s τ³-Voice and Cekura, all dated October 29.

On Sierra’s multilingual full-duplex τ³-Voice benchmark, Phonic reports a score of 83.2 for Phonic-1, ahead of GPT-Live-1 at 79.9 and Gemini 3.8 Live at 68.5. The model is trained on more than 50 languages and, according to the company, can switch language mid-conversation while keeping the same speaker identity.

Phonic is unusually direct about where the model falls short. In blind human evaluations run by an unnamed third party, judges preferred Phonic-1 over other models “the majority of the time.” But when it was compared with real people, Phonic says “humans win ~60% of the time,” and calls naturalness a core research direction for future models.

Where it runs, and who is using it

Phonic says the model already handles calls on six continents for customers including Ashby, Maven AGI, Rho and an unnamed Fortune 50 IT infrastructure company, across support, healthcare, recruiting and logistics. Ashby co-founder and VP of engineering Abhik Pramanik is quoted as saying Phonic “powers our AI interviews in 25 countries.”

On deployment, Phonic’s site lists a multi-tenant Phonic Cloud by default, single-tenant hosting in a chosen region, VPC peering and an enterprise self-host option. It says it is SOC 2 Type II certified, offers HIPAA business associate agreements and connects through webhooks, a held-open websocket or LiveKit. Pricing is a single per-minute rate, though the company does not publish the figure on its homepage.

Phonic-1 is a closed, hosted model, and Phonic has not released a model card, parameter count or technical report. The new money, the company says, goes toward scaling speech-to-speech in production, with hiring across research, engineering, product and business roles.