Phonic-1: A New Speech-to-Speech Model Claims 600 ms and Tops One Index, Claimed vs Verified

Phonic, a voice AI startup, has launched Phonic-1, which it calls "the most advanced speech-to-speech model for agentic tasks," and announced a $27.5M Series A led by First Round with Lux Capital, Felicis and Definition, according to Phonic's announcement. The pitch is a phone agent that answers in about 600 milliseconds, hears tone and hesitation rather than only words, and can call tools while the caller is still talking. The independent Artificial Analysis speech-to-speech leaderboard lists Phonic 1 (High) at an index of 87.1, the top number on that page when we read it on October 29.

A launch with a benchmark lead is exactly where claims deserve a check. Below: what Phonic says, what the independent data shows, what the product costs, and what stays unverified. This follows our earlier look at a similar claim in Rime Prism's 300 ms speech-to-speech pitch versus its published latency.

TL;DR: claimed versus verified

What did Phonic announce?

Phonic's homepage describes four selling points. It says callers speak naturally because Phonic "hears tone and hesitation, not just words." It says Phonic responds in 600 ms "while other agents take a full second or more." It says it serves callers in 50+ languages. And it says "Phonic-1 is the only speech-to-speech model that can call tools and take action while the caller is speaking."

It also states that Phonic "uses frontier models as supervisor agents to assist in tool calling and reasoning." That sentence is important. A speech-to-speech model handles the audio path end to end, but the heavy reasoning and tool decisions are partly delegated to other models. The company says it owns the full audio path and routes speech into reasoning, so builders get "one model, one price per minute" instead of stitching speech-to-text, a language model and text-to-speech together.

Integration is through webhooks, a held-open websocket, a no-code builder or LiveKit, with an SDK in TypeScript and Python. Enterprise options include single-tenant deployment, VPC peering and a self-hosted tier, plus SOC 2 Type II, HIPAA with a BAA, a stated 99.99% uptime SLA and SAML or OIDC single sign-on. Those are Phonic's statements; we did not audit them.

What does the independent data say?

Artificial Analysis publishes a speech-to-speech index that averages four things: speech reasoning (Big Bench Audio), agentic performance (tau-Voice), arena preference and task success rate. As of our read on October 29, the page lists these figures for Phonic 1 (High):

Source: Artificial Analysis, read October 29, 2026. It is a live page and numbers can change. The conversational dynamics score for Phonic is shown as unavailable.

Three readings are fair. First, the 600 ms claim lines up with the 0.61 second time to first audio. Second, the index lead over Grok Voice 3.0 is 0.2 points, which is within what a methodology tweak could change, so "tops the index" is accurate but not decisive. Third, Phonic does not lead every component: Grok Voice 3.0 is ahead on arena preference and task success rate, while Phonic is ahead on the agentic tau-Voice score and speed.

Our coverage of Grok Voice 3.0 and its 86.9 index and of ChatGPT's GPT-Live-1.1 rollout give the context for those rivals.

What does it cost?

Two speech bubbles passing a bead between them, picturing turn-taking in a voice call

Phonic's pricing page says "Pay by the minute. No LLM pass-throughs. No per-token charges."

Both self-serve plans carry $250 per month of free credits for accounts that sign up by November 30. At $0.07 per minute, a ten-minute call costs about 70 cents; a thousand such calls cost about $700. Compare that with the $4.20 per hour of input audio Artificial Analysis computes on its benchmark subset, which is token-based and not directly comparable to a per-minute plan.

A flat per-minute price is easier to budget than token billing, and Phonic's own engineering post on concurrency limits argues that capacity caps are a hidden cost in voice AI. The trade-off is lock-in to one provider's stack and no visibility into which underlying models handle your reasoning.

What is still unverified?

• Tool calls while the caller speaks. Phonic's headline differentiator has no independent test we could find, and "only" is a strong word given how fast rivals iterate.
• Real-phone latency. Time to first audio on a benchmark is not the same as a caller's experience over a telephone network. Test on your carrier and region.
• Emotion handling. "Hears tone and hesitation" is a product claim. The index components do not measure it directly.
• Conversational dynamics. The leaderboard shows no score for Phonic on full-duplex behaviors like interruption and backchannels, which matter for natural calls.
• Customer results. Phonic's blog lists case studies, such as call completion rising from 35% to 90% at one customer, but these are vendor-written and we did not verify them.
• Funding terms. The $27.5M figure comes from the company; we did not see an independent filing.

How do you read these benchmarks?

The four components behind the index measure different things, and it helps to know what each can and cannot tell you.

• Speech reasoning (Big Bench Audio). One thousand audio questions adapted from Big Bench Hard, covering formal fallacies, navigation, object counting and web of lies. Several models, including Phonic, Grok Voice 3.0 and Qwen's realtime model, score between 98% and 99%, so this component no longer separates the leaders. It mostly shows a model is not broken.
• Agentic performance (tau-Voice). A simulated caller phones a support agent with a problem such as a flight change or a disputed charge, and the result is judged on the final database state. This is the closest public test to what Phonic markets, and it is where Phonic's 79.9% stands out. It still uses simulated callers and scripted domains, so it is not your business.
• Arena preference. Humans pick between two unnamed models after the same scenario, such as booking a dental appointment. This captures feel, and Phonic's Elo of 1012 sits below Grok Voice 3.0 and GPT-Live-1 (Astra), a reminder that a high index does not mean people prefer the voice.
• Task success rate. The share of conversations that end in the correct tool call. Phonic's 92.3% is strong but behind Grok Voice 3.0's 94.8% and Gemini 3.8 Live's 93.2%.

Put differently, Phonic looks best where the test is a clean, scripted, tool-using call and merely competitive where humans judge. That matches its positioning at enterprises that run high-volume operational calls, such as trucking dispatch, screening and intake, which are also the customers in its published case studies.

Latency deserves its own caution. Artificial Analysis measures time to first audio on a benchmark with the model's own API, not over a telephone network. A real call adds carrier delay, your tool's response time and any supervisor-model round trip. If Phonic hands a hard question to a frontier model in the background, a caller may still hear a pause, even if the first audio of a filler phrase arrives quickly. That is why testing a slow tool call is the most revealing experiment you can run.

How should builders evaluate it?

1. Run your own calls. Use a script of your hardest scenarios, including interruptions, accents, background noise and a tool call that takes two seconds, and measure end-to-end time on a real phone line.
2. Check the supervisor layer. Ask which frontier models handle reasoning, where data flows, and how that affects HIPAA and retention commitments.
3. Compare at your volume. Per-minute pricing wins at predictable usage; token pricing from other providers can win for short or sparse calls.
4. Test failure modes. Voice agents that act on real systems can be talked into mistakes. Put limits on what tools can do; AgentBeam is the agent security platform from the explainx.ai team for stopping AI agents before they take dangerous actions.
5. Re-read the leaderboard. It is live, and a 0.2 point lead can flip.

Where does this fit in the voice AI market?

Voice is crowded and moving quickly. Deepgram bought Daily and kept Pipecat open source, AssemblyAI shipped a text-to-speech model at $15 per million characters, and ElevenLabs offers Eleven Realtime v1 Mini and Max. Phonic's bet is that a specialist that owns the audio path can beat generalist labs on agentic phone work. The leaderboard supports that for now; sustained leads are harder.

Bottom line

Phonic-1 launches with credible independent numbers behind its speed and agentic claims and a clear per-minute price, which is more than many launches offer. The distinctive claim, tool calls during speech, and the emotional-awareness claim remain the company's word. If you build phone agents, trial it against Grok Voice and GPT-Live on your own call scripts before moving anything into production. We will update this post if independent tests of the tool-call claim appear.

Related reading

• Rime Prism: 300 ms claim versus 570 ms published
• Grok Voice 3.0: 86.9 speech-to-speech index
• ChatGPT GPT-Live-1.1: voices, languages and limits by plan
• Deepgram acquires Daily; Pipecat stays open source
• AssemblyAI TTS-1 for voice agents
• ElevenLabs Eleven Realtime v1 Mini and Max

Prices, leaderboard figures and claims are as of October 29, 2026 and may change.