Latest in AI
Phonic-1 Released: Speech-to-Speech Pricing, Benchmarks and Tool Calls
Phonic's new voice model can start tool calls while a caller is speaking, making permissions and corrections as important as response speed.
Phonic released Phonic-1 on October 29, a speech-to-speech model for voice agents that use tools and complete tasks. The company says it can initiate tool calls while a caller is still speaking. It also announced a $27.5 million Series A led by First Round.
The interesting change is where work happens during a conversation. If an agent can look something up before the caller finishes, it can reduce the awkward pause afterward. But starting work earlier also makes it more important to distinguish a harmless lookup from an action that needs confirmation.
What Phonic-1 changes in a voice agent
A cascaded voice system typically connects speech recognition, a language model and speech synthesis. A speech-to-speech model handles audio directly instead of relying on that sequence as its core interface.
In its launch announcement, Phonic describes tool use during incoming speech and says Phonic-1 is available immediately. It also says the model was trained on more than 50 languages and can switch languages while preserving the speaker's voice identity. These are vendor descriptions, not a guarantee that every language and workflow performs equally well.
The tool-use mechanism suggests a useful distinction for application builders. Looking up a flight while someone explains a problem can be helpful. Changing the booking before they finish explaining would create a different risk. The product still needs explicit rules about which operations may run speculatively and which require a completed instruction or confirmation.
Phonic-1 pricing
Phonic's published pricing uses conversation minutes rather than separate model-token charges.
| Plan | Listed price | Monthly minimum | Concurrent calls |
|---|---|---|---|
| Starter | $0.07 per minute | None | 10 |
| Pro | $0.065 per minute | $200 | 50 |
| Enterprise | Custom | Custom | Custom |
The company says there are no separate LLM pass-through or per-token charges. At the Starter rate, 1,000 conversation minutes would be $70 before any other applicable commercial conditions. That is simple arithmetic, not a tested invoice.
A minute-based bill is easier to forecast than several token meters, but it still rewards an efficient conversation. Repeated questions and failed tool calls can extend a call even when the voice sounds natural.
What the independent benchmarks show
On the Artificial Analysis speech-to-speech leaderboard, Phonic 1 in its High configuration scores 87.1 on the Speech to Speech Index, compared with 86.9 for Grok Voice 3.0. Phonic's listed agentic-performance score is 79.9%, and its time to first audio is 0.61 seconds.
The narrow index lead should not be read as superiority on every measure. Time to first audio measures when output begins, not how long a booking or support case takes to finish. Configuration and evaluation design also matter when comparing the figures.
A separate Cekura evaluation reports 90.2% reliability for Phonic 1 using pass³: the share of scenarios that pass every required check across three retained runs. Its median response time is 1.63 seconds; Grok Voice 3.0 is faster on that measure at 1.34 seconds.
Those latency numbers are not contradictory. They come from different tests and measure different aspects of the interaction. Cekura also records remaining errors, including occasional mistakes in saved handoff or routing details. A strong aggregate score does not make the resulting business record automatically correct.
The useful test is a correction halfway through a request
For a deployment, test a caller who changes their mind after a lookup has already begun. The system should keep useful retrieved information while abandoning an action based on the earlier instruction. That is a harder requirement than simply replying quickly.
Phonic-1 makes that evaluation more relevant. As voice agents move work into the middle of a sentence, the next improvement needs to be measured in correctly completed tasks, with the caller's final intent preserved.
I’m Paras. I work at the intersection of AI, engineering and distribution.
I help technical teams turn model launches and complex product workflows into clear stories developers and founders want to read and share.