Phonic-1: Realtime Voice Agents, Pricing and Setup

Phonic-1 is designed for an interactive conversation: a caller speaks, the agent responds aloud, and the application can expose tools for information or actions. A conventional speech pipeline combines transcription, a text model and speech synthesis. Phonic supplies the speech-to-speech model as one hosted service, so those three model choices are not separate integration steps.

The publisher says Phonic can begin calling tools while the caller is still speaking and can switch among more than 50 trained languages while keeping the same speaker identity. These features are useful when a caller adds details gradually or corrects an earlier request. They still need testing with your tool response times, languages and audio conditions.

Use phonic-1 when configuring this release. The API schemas also list phonic-0.5 and phonic-1.1; those enum entries do not establish that every account can use every version. Set the intended model on the agent instead of relying on a rolling default when comparing results or costs.

Phonic’s current pricing page bills per conversation minute. Starter lists $0.07 per minute without a monthly minimum; Pro lists $0.065 per minute with a $200 monthly minimum. The page describes a unified model price without additional LLM token pass-through charges. A lower per-minute rate only reduces the month’s bill once usage justifies the minimum.

At an assumed 1,000 billed minutes per month, Starter costs $70. The same usage on Pro would calculate to $65 at the usage rate but remains subject to the $200 minimum. At 10,000 minutes, Starter is $700 and Pro is $650. These examples use the published rate and minimum, before taxes and separate services.

Add the telephony and transport costs of the route you choose. The custom-number guide uses externally purchased numbers and SIP trunks. A LiveKit integration can also have its own infrastructure or Cloud usage bill. Phonic’s model-minute price alone does not establish that those costs are included.

For a support workflow, divide the full voice, tool and transport bill by completed calls. Track transfers, failed lookups, repeat calls and human review separately. Ten minutes of conversation can resolve a case or leave it unfinished; the per-minute tariff does not measure that difference.

Start in Phonic’s agent builder when you want to test a system prompt, voice and tools before writing an integration. Create the agent, select the intended model, attach the tools and try a bounded call. Use a lookup or other reversible operation for the first test so you can inspect the arguments and response.

With webhooks, Phonic manages the session and your HTTP endpoints provide configuration and receive call results. A WebSocket connection gives your application control over a realtime audio connection and custom tool paths. Choose that route when voice belongs inside your own application or tools need a private connection.

Use the LiveKit plugin when a LiveKit AgentSession already owns the surrounding voice workflow. Phonic’s guide notes that this adds an intermediate hop and exposes a subset of its configuration. It does not quantify a universal latency penalty. Compare both routes with the same agent, voice, tool server and caller region before deciding.

Create a Phonic agent with phonic-1 selected and a short instruction for one task, such as looking up an appointment. Give each tool a clear input schema and return an explicit success or failure. Store PHONIC_API_KEY on the server that hosts your integration. Follow the chosen deployment guide for its session and audio connection.

In LiveKit, install the Phonic plugin and use phonic.realtime.RealtimeModel in AgentSession. The plugin’s phonic_agent setting selects an existing Phonic agent. Configure the model there; a voice name alone does not pin the model. LiveKit’s agent instructions take precedence over the system prompt saved on that Phonic agent.

Check a full interaction: greet, look up a value, correct one detail, interrupt a response and end the call. Inspect the tool arguments and final saved record. A plausible spoken answer is insufficient if the agent saved the wrong date, called the wrong tool or failed to finish the transaction.

Cekura’s speech-to-speech benchmark tests whole voice-agent configurations using the same Pipecat agent, prompt, tools and connection. Its October 22 results identify Phonic 1 with the recorded model alias phonic_v1, high intelligence and the Sabrina voice. The suite contains 82 scenarios, each run three times: 246 calls in total.

Phonic passed all three runs on 74 of the 82 scenarios, or 90.2% of scenarios. Across individual calls, success was 95.1% and data accuracy 97.5%. Median caller waiting time for a response was 1.63 seconds, with a reported 1.81-second p90. Those are results for the complete test configuration, rather than a latency guarantee for another transport or application.

Cekura also reports failures, including ending a call without acknowledging the caller’s request to call later and occasionally saving incorrect transfer notes or routing details. Its interruption score is 4.88 out of five. Review the task-level results when interruption handling or exact data capture matters more than an aggregate ranking.

The benchmark provider publishes reference code, while parts of the judging setup are not public. Treat it as third-party evidence with a defined harness. It does not supply a production customer’s invoice or a controlled comparison between Phonic direct integration and LiveKit.

Phonic is worth a trial when conversation handling and tool-driven work need to operate together, and one hosted speech model fits the deployment. A modular transcription/text/TTS stack offers more component-level choice. Compare the complete accepted workflow rather than assuming either architecture wins from a demo.

For a prerecorded transcript, Nemotron Transcribe serves a different task. For generated Foley and atmosphere, SFX v3 creates assets rather than live conversation. GPT-Live is another realtime model to compare on the same call scripts, languages and tools.