Most phone agents still work like a slow relay race. You speak, the system writes down what you said, a language model thinks about the text, a voice model reads the answer back, and only then does anything actually happen. Phonic-1 is built to collapse that whole chain. Launched on October 29, 2026 by the voice AI company Phonic, it is a speech-to-speech model that understands spoken words directly, without turning them into text first, and it can start calling tools while you are still talking. The result, Phonic says, is a phone agent that answers in about 600 milliseconds and has its answer ready the moment you stop speaking.
In this guide, we explain what Phonic-1 actually is, how its speech-to-speech design differs from the voice agents most businesses use today, what the independent benchmark numbers say (and what they do not), how much Phonic-1 costs, and where its honest limits sit. We also clear up a name trap that could easily confuse buyers: Phonic-1 is not the same thing as Ollama Phonon, a similarly named but completely different product. Every benchmark figure in this article is labelled as either company-reported or independently measured, because the two are not the same thing, and a buying decision should never mix them up.

Table of Contents
- What Is Phonic-1?
- How Phonic-1 Works: Speech In, Speech Out, No Text Step
- Phonic-1 Benchmarks: 87.1 Index, 600ms, and 79.9% Task Success
- Phonic-1 vs Grok Voice, GPT-Live, and Gemini Live
- Phonic-1 Pricing: $0.07 Per Minute and $250 in Free Credits
- Who Should Use Phonic-1?
- Phonic-1 vs ElevenLabs and Mistral Voxtral 3
- Phonic-1 vs Phonon: Do Not Confuse the Names
- Phonic-1 for Developers: API, Tools, and First Build
- Limitations and Honest Caveats
- How to Get Started With Phonic-1
- Verdict: Is Phonic-1 Worth Trying?
- Frequently Asked Questions About Phonic-1
What Is Phonic-1?
Phonic-1 is a speech-to-speech AI model for agentic phone tasks, built by Phonic, a voice AI company that says it has focused exclusively on speech-to-speech technology since it was founded more than two years ago. Unlike a traditional voice assistant, which chains together three separate systems (speech recognition, a text language model, and text-to-speech), Phonic-1 takes audio in and produces audio out in a single model. That design choice is the whole point of the product: because nothing is converted to text in the middle, the model can hear tone, hesitation, and interruptions, and it can begin acting before a sentence is finished.
Phonic describes Phonic-1 as “the most advanced speech-to-speech model for agentic tasks.” That is marketing language, and it should be read as such. What can be checked independently is the model’s position on the Artificial Analysis Speech-to-Speech Index, where Phonic 1 (High) held the top score of 87.1 when the leaderboard was read on October 29, 2026, just 0.2 points ahead of Grok Voice 3.0.
The launch was paired with a $27.5 million Series A funding round led by First Round, with Lux Capital, Felicis, and Definition participating, bringing Phonic’s total funding to $31.5 million. Funding does not make a model good, but it does signal that investors expect speech-to-speech agents to become a real business category, not a demo trick.
Phonic says its model already powers calls across six continents for customers including Ashby, Maven AGI, and Rho, plus a Fortune 50 IT infrastructure company, and that its AI interviews run in 25 countries. Those are company-reported customer claims, not independently audited figures, but they matter for one practical reason: Phonic-1 is positioned as production software for real phone lines, not a research preview. It is available through the Phonic API today, and Phonic offers a live demo phone line so prospective buyers can hear the agent before paying anything.
How Phonic-1 Works: Speech In, Speech Out, No Text Step
The traditional voice-agent pipeline has a built-in delay problem. Each handoff between systems adds latency: the transcriber must decide you have finished speaking, the language model must generate a full text reply, and the voice synthesizer must render that reply as audio. On a phone call, those delays feel like awkward pauses, and callers fill the silence by repeating themselves or hanging up. Phonic-1 removes the handoffs by keeping everything in the audio domain. The model listens to speech, reasons over it, and speaks its reply, all as one continuous process.
That architecture enables the feature Phonic is marketing hardest: tool calls while the caller is still talking. In Phonic’s launch example, Phonic-1 called five external tools during a single caller utterance and delivered its answer 1.1 seconds after the speaker finished. Think about what that means in practice.
If a caller says, “I need to change my delivery to Thursday and check whether the large size is in stock,” a conventional agent waits for the sentence to end, then looks up the order, then checks inventory, then answers. Phonic-1 can start the order lookup and the inventory check while the words are still coming, so the answer is ready almost immediately. Phonic claims it is the only model that can act while the caller is still talking; that specific “only” claim has not been independently tested, so treat it as a company claim rather than an established fact.
Phonic-1 and Natural Conversation: Tone, Interruptions, and Noisy Calls
Phonic says the model was post-trained for natural conversation, meaning it is designed to understand tone and hesitation, handle turn-taking, and keep working in noisy environments. This matters more than it sounds. Real phone calls are messy: people interrupt, talk over each other, pause mid-sentence to think, and call from cars and warehouses. A model that only handles clean, one-speaker-at-a-time audio will fail in exactly the situations where a business most needs it. Speech-to-speech models have a structural advantage here because tone and pacing are part of the input, not information thrown away by a transcript. A sigh, a rushed sentence, or a hesitant “well…” all carry meaning that text transcription discards.
It is trained on more than 50 languages and can switch languages in the middle of a conversation while keeping the same speaker identity. For a US business, the practical value is obvious: a caller who starts in English and switches to Spanish mid-call does not have to be transferred, re-authenticated, or handed to a different agent. The same voice continues, in the new language. That is a genuine differentiator against older voice systems, which typically treat a language switch as a new session.

Phonic-1 Benchmarks: 87.1 Index, 600ms, and 79.9% Task Success
Benchmarks are where the story gets interesting, because this launch has both company claims and independent measurements, and they do not all say the same thing. Here is the picture, with each number labelled by its source.
On the independent side, Artificial Analysis lists Phonic 1 (High) at 87.1 on its Speech-to-Speech Index, the top score on that page on October 29, 2026. The index blends four equally weighted components: the Big Bench Audio reasoning benchmark, the τ-Voice customer service benchmark, Speech Agent Arena preference ratings, and task success rate.
Phonic 1 (High) recorded a 79.9% task success rate on that test set, a time to first audio of 0.61 seconds, and a cost of $4.20 per hour of input audio on a 40-question subset of Big Bench Audio. That 0.61-second figure lines up closely with Phonic’s own “about 600 milliseconds” claim, which is a good sign: when a company’s headline speed number matches an independent measurement, the claim is much easier to trust.
There are two honest qualifications. First, Phonic 1 was the fourth fastest model on the leaderboard, not the fastest, though it was the fastest of the top five models by index score, which is the more meaningful comparison for buyers who care about both intelligence and speed. Second, the $4.20-per-hour benchmark figure excludes cache discounts and tool-call costs, so it is not a complete bill. Your real cost per hour will depend on how many tools your agent calls and how much audio your callers actually produce.
On the company-reported side, Phonic cites a score of 83.2 on Sierra’s multilingual full-duplex τ³-Voice benchmark, ahead of GPT-Live-1 at 79.9 and Gemini 3.8 Live at 68.5. That is a Phonic-selected benchmark reported by Phonic, not an independent re-run, so read it as supporting evidence rather than proof. Phonic also says the model sits alone on the Pareto frontier of intelligence and latency on end-to-end full-duplex benchmarks. Again, that framing comes from the vendor. The independent Artificial Analysis numbers are the ones to anchor on, and they are strong enough on their own without the marketing gloss.
Phonic-1 and the Human Preference Gap
One number deserves special attention because it cuts against the hype: human listeners still preferred real people about 60% of the time in preference testing reported around the launch. That is not a failure of Phonic-1 specifically; it reflects where the whole category sits in late 2026. Voice agents have become fast and capable, but callers can still tell the difference, and many still prefer a human when the conversation matters.
For a business, the honest takeaway is that Phonic-1 is best deployed where speed and availability matter most (after-hours support, appointment changes, order tracking, screening interviews) and where a human handoff is available for the conversations that need one. The goal is not to replace every human call; it is to stop losing the calls that currently go unanswered.
Phonic-1 vs Grok Voice, GPT-Live, and Gemini Live
The lead on the Artificial Analysis index is real but narrow: 87.1 against 86.9 for Grok Voice 3.0. A 0.2-point gap is not a commanding victory, and it would be a mistake to choose a voice platform on a single leaderboard number. What the comparison does show is which trade-offs each model makes. It is the fastest of the top five models by index score and the cheapest of the leaders on benchmark input-audio cost after Google’s Gemini 3.8 Live Extended Thinking (High). Grok Voice 3.0 and the Sol and Astra versions of GPT-Live-1 cost more per hour of benchmark input audio than Phonic 1. For a business running thousands of minutes a month, that per-hour difference compounds quickly.
Here is how the leading speech-to-speech options compare on the numbers available at launch:
| Model | Speech-to-Speech Index | Time to First Audio | Task Success | Benchmark Cost per Hour of Input Audio |
|---|---|---|---|---|
| Phonic 1 (High) | 87.1 | 0.61s | 79.9% | $4.20 |
| Grok Voice 3.0 | 86.9 | Not disclosed in the same table | Not disclosed in the same table | Higher than Phonic 1 |
| GPT-Live-1 (Sol / Astra) | Below Phonic 1 | Not disclosed in the same table | Not disclosed in the same table | Higher than Phonic 1 |
| Gemini 3.8 Live Extended Thinking (High) | Below Phonic 1 | Not disclosed in the same table | Not disclosed in the same table | Lower than Phonic 1 |
The table above mixes independently measured figures (Phonic 1’s row) with directional comparisons reported from the same leaderboard, and we have marked the gaps honestly rather than inventing numbers for rivals. The point for buyers is not that the model wins every column; it is that Phonic-1 is the only model in this group that combines a top index score with the fastest time to first audio among the leaders and a mid-call tool-calling design. If your use case is a phone agent that must act, not just talk, that combination is what you are actually paying for.
Phonic-1 Pricing: $0.07 Per Minute and $250 in Free Credits
Phonic’s pricing is refreshingly simple for the voice AI market, where many rivals hide behind token math and platform fees. There are two published tiers:
- Starter: $0.07 per minute. The entry tier for trying Phonic-1 and running smaller call volumes.
- Pro: $0.065 per minute. A lower per-minute rate for higher-volume or production use.
At $0.07 per minute, an hour of talk time costs $4.20, which matches the benchmark input-audio cost quoted above and gives a useful rule of thumb: budget roughly $4.20 per agent-hour before tool-call and telephony costs. A 10-minute support call costs about $0.70 in model time; a 30-minute screening interview costs about $2.10. Those numbers make Phonic-1 easy to compare against a human agent’s loaded cost, and for most US businesses the model cost is a small fraction of what a staffed phone line costs.
Phonic is also offering $250 in free credits for sign-ups by November 30, 2026. At the Starter rate, that covers roughly 3,570 minutes, or nearly 60 hours of calls, which is a generous trial by voice AI standards. Two things to check before you rely on that figure: first, whether the credit covers tool-call costs and telephony charges or only model minutes; second, what happens to your agents and call history if you do not convert to a paid tier after the credits run out.
Neither detail was in the launch announcement, so confirm both with Phonic before building a production workflow on the free credits. Phonic also advertises a 99.99% uptime SLA, which matters if your agent answers a sales or support line where downtime means lost revenue.

Who Should Use Phonic-1?
Phonic is aiming Phonic-1 at phone-first work, and the three launch use cases tell you a lot about where it fits:
- Customer support. Order status, appointment changes, account questions, and first-line troubleshooting, especially after hours when staffing a phone line is expensive. The mid-call tool-calling design is built for exactly this: look up the order while the customer is still explaining the problem.
- Job interviews. Structured screening interviews at scale, already used in 25 countries according to Phonic. A voice agent that hears hesitation and tone can run a fairer, more consistent first screen than a rigid questionnaire, and it never gets tired on the fiftieth call of the day.
- Logistics. Delivery coordination, driver check-ins, and dispatch calls, where the caller is often in a noisy environment and the agent must act on what it hears immediately rather than after a long exchange.
Beyond those three, the model makes sense for any US business where missed calls mean missed money: clinics confirming appointments, home-service companies booking jobs, restaurants handling reservations, and e-commerce brands answering “where is my order” at midnight. It makes less sense for conversations where empathy and judgment are the product itself, such as debt counselling, medical triage, or complaint escalation. There, the right design is a Phonic-1 agent that qualifies and routes, with a human taking over the moment the conversation turns sensitive. That hybrid pattern is where voice agents earn their keep today.
Phonic-1 vs ElevenLabs and Mistral Voxtral 3
Voice AI buyers in late 2026 face a crowded field, and Phonic-1 is not the only new agent stack. Two recent launches frame the choice well. Mistral’s Voxtral 3 and ElevenLabs’ Eleven Realtime v1 both shipped in October 2026 as full voice-agent stacks, and our comparison of the two is a useful companion read. The key difference is architectural: Voxtral 3 and Eleven Realtime are speech-to-speech systems too, but the headline bet is acting mid-utterance, calling tools before the caller finishes.
If your agent mostly answers questions from knowledge it already has, any of these platforms can do the job, and the choice comes down to price, language coverage, and ecosystem. If your agent must check systems, book things, and change records during the call, its design is aimed squarely at that job.
Price is the other separator. ElevenLabs and Mistral price their agent stacks differently, with platform and model fees that vary by configuration, while the flat per-minute rate is easy to forecast. For a small business that wants a predictable phone bill, that simplicity is a real selling point. For a large enterprise already committed to an ElevenLabs or Mistral ecosystem, switching costs may outweigh a per-minute saving. There is no universal winner here, only a best fit for a given call volume, tool complexity, and existing stack. If you are still choosing between voice platforms generally, our best AI voice generator guide covers the wider field, from narration and dubbing tools to live agents.
Phonic-1 vs Phonon: Do Not Confuse the Names
A warning that could save you a wasted afternoon: Phonic-1 is not Phonon. Ollama Phonon is a text-to-speech model family from a different company, Gradium, that runs locally on your Mac and turns written text into spoken audio. It is a free, local narration tool. Phonic-1 is a speech-to-speech agent model from the company Phonic that holds live phone conversations and calls tools. The names sound almost identical, the products do completely different jobs, and search results will happily mix them up. If you are shopping for a phone agent, you want Phonic-1. If you want free local text-to-speech on your own machine, you want Phonon. Do not buy, install, or benchmark one while thinking it is the other.

Phonic-1 for Developers: API, Tools, and First Build
For developers, the interesting part of this launch is not the leaderboard score but the programming model it implies. A speech-to-speech agent that calls tools mid-utterance changes how you design a call flow. In a traditional pipeline, your code waits for a completed transcript, runs a function, and feeds the result back into the next turn. With this model, your tools need to be ready to fire early, return partial or streaming results, and handle the case where the caller changes direction halfway through a sentence. That means designing tools that are idempotent (safe to call twice), fast to start, and able to return a useful partial answer rather than failing when the full picture is not ready yet.
The practical checklist for a first build starts with tool design. Keep each tool narrow: one tool that checks an order, one that checks stock, one that books an appointment. Narrow tools are easier to test, easier to permission, and easier to explain to a caller when something goes wrong. Give each tool a clear failure message in plain language, because the agent will speak that failure to the caller. A tool that returns a raw database error will produce a confusing phone call; a tool that returns “I could not find an order with that number, would you like to check the number?” produces a recoverable one.
Next, decide what the agent may never do without confirmation. Moving money, cancelling a booking, changing a medical appointment, or sharing account details should require an explicit spoken confirmation, even if the agent is technically able to do it in one step. The speed of this model makes it tempting to let it act on everything immediately, but speed without guardrails is how a misunderstood sentence becomes a cancelled order. Build a confirmation step for every irreversible action, and log every tool call with the audio timestamp so disputes can be reviewed.
Finally, test with real noise. Demo calls in a quiet room flatter every voice model. Test from a car, a warehouse floor, and a busy street, with callers who interrupt, mumble, and change their minds mid-sentence. That is the environment the model claims to handle, and it is the environment your customers actually call from. If the agent holds up there, the benchmark numbers stop mattering, because you have measured the thing that counts: completed calls.
Limitations and Honest Caveats
This is a strong launch, but a 2,500-word guide would be dishonest if it only repeated the press release. Here are the limits that matter:
- The “only model” claim is unverified. Phonic says Phonic-1 is the only model that acts while the caller is still talking. No independent test of that specific claim has been published. Treat it as a design claim with a compelling demo, not an established fact.
- Benchmark speed is not phone-call latency. The 0.61-second time to first audio comes from a controlled benchmark, not from a live phone network with real-world jitter, carrier routing, and telephony hops. Expect real calls to be a little slower.
- Humans still win on preference. Listeners preferred real people about 60% of the time in preference testing around launch. Deploy Phonic-1 where speed and coverage matter, and keep humans in reach for the rest.
- The 0.2-point leaderboard lead is thin. Grok Voice 3.0 sits at 86.9 against Phonic-1’s 87.1. Leaderboard positions shift with every update, so do not build a multi-year vendor decision on a single decimal.
- Pricing gaps remain. The published per-minute rates do not spell out telephony charges, tool-call costs, or what happens when free credits expire. Confirm all three before going live.
- Company benchmarks are company benchmarks. The Sierra τ³-Voice score of 83.2 and the “Pareto frontier” framing come from Phonic. The independent Artificial Analysis figures are the anchor; the rest is supporting color.
How to Get Started With Phonic-1
Getting started is straightforward, and the $250 credit offer makes a real trial possible before you spend anything:
- Try the live demo first. Call Phonic’s public demo line and have a real conversation with a Phonic-1 agent. Judge it the way your customers will: does it interrupt gracefully, does it understand you in a noisy room, and does the answer arrive fast enough to feel natural?
- Sign up before November 30, 2026 to claim the $250 in free credits, and confirm in writing whether tool-call and telephony costs are included.
- Build one narrow agent first. Pick a single high-volume, low-risk flow, such as order-status checks or appointment rescheduling, and connect it to one or two tools. Measure resolution rate and caller satisfaction before expanding.
- Design the human handoff on day one. Decide exactly which phrases, topics, or caller emotions trigger a transfer to a person, and test that path as carefully as the happy path.
- Watch the bill for a week. Track real cost per completed call, including tools and telephony, and compare it with your current cost per call before committing to the Pro tier at $0.065 per minute.

Verdict: Is Phonic-1 Worth Trying?
This launch earns its attention. A top independent index score, a sub-second time to first audio that matches the company’s own claim, a genuinely different design idea in mid-utterance tool calling, and a simple $0.07-per-minute price with $250 in free credits add up to one of the most trial-friendly voice agent launches of 2026. For phone-first businesses, especially support, scheduling, interviews, and logistics, it is absolutely worth a test call and a one-flow pilot.
It is not, however, a reason to fire your support team or to sign a long contract on launch week. The leaderboard lead is thin, the boldest claim is unverified, human callers still prefer humans most of the time, and the full cost picture depends on telephony and tool charges that are not in the headline price.
The smart move is the boring one: run the demo, spend the free credits on a single real workflow, measure cost per resolved call, and then decide. On the evidence available at launch, it looks like a serious contender in the voice agent race, and its speech-to-speech approach is the direction the whole category is heading. Whether it stays at the top of the index is a question for the next benchmark update, not for a press release.
Frequently Asked Questions About Phonic-1
What is Phonic-1?
Phonic-1 is a speech-to-speech AI model from Phonic, launched October 29, 2026, that holds live phone conversations and calls tools while the caller is still talking, without converting speech to text in the middle.
How much does Phonic-1 cost?
The Starter tier costs $0.07 per minute and $0.065 per minute on Pro. New sign-ups by November 30, 2026 receive $250 in free credits. Confirm whether telephony and tool-call charges are included before budgeting.
How fast is Phonic-1?
Phonic claims responses in about 600 milliseconds. Artificial Analysis independently measured a time to first audio of 0.61 seconds for Phonic 1 (High), and a demo showed an answer delivered 1.1 seconds after the speaker finished, after five tool calls during the utterance.
Is Phonic-1 the same as Ollama Phonon?
No. Phonic-1 is a speech-to-speech phone agent from Phonic. Ollama Phonon is a text-to-speech model family from Gradium that runs locally. Different companies, different products, similar names.
Which languages does Phonic-1 support?
Phonic-1 is trained on more than 50 languages and can switch languages mid-conversation while keeping the same speaker identity, according to Phonic.
Is Phonic-1 better than Grok Voice 3.0?
On the Artificial Analysis Speech-to-Speech Index read on October 29, 2026, Phonic 1 (High) scored 87.1 against Grok Voice 3.0 at 86.9, a narrow lead. Phonic-1 was also the fastest of the top five models on that page. The gap is small, so choose on your use case and total cost, not the leaderboard alone.
Sources: Phonic’s launch announcement “Introducing Phonic-1 and our $27.5M Series A” (October 29, 2026); Artificial Analysis Speech-to-Speech Index (October 2026); independent launch coverage of the Phonic-1 announcement. Company-reported figures are labelled as such in the article.
