Phonic 1 Debuts at #1 on the Speech Index. The 0.2-Point Margin Hides the Real Lead.

Phonic 1 is now the top-ranked speech-to-speech model on Artificial Analysis's Speech to Speech Index at 87.1, a tenth of a point race it won by edging xAI's Grok Voice 3.0 at 86.9. A 0.2-point margin would normally be noise, and on this index it very nearly is: the two models are not winning the same events, so the composite tie hides a much larger split underneath it. On the agentic task benchmark that feeds a quarter of that score, Phonic 1 lands at 79.9% against 74.5% for the best OpenAI configuration.

Phonic, a roughly ten-person spinout of MIT and Stanford researchers, introduced the model on October 29, 2026 alongside a $27.5 million Series A led by First Round, bringing its total funding to $31.5 million. The company spells the model Phonic-1 in its own launch post and documentation; Artificial Analysis lists it as Phonic 1 (High), where the parenthetical is an API setting rather than a product name. Both spellings point at the same endpoint and the same $0.07-per-minute list price.

What Phonic-1 is, in the vendor's own terms

It is an end-to-end speech model, not a cascade of transcribe-then-think-then-speak. Audio goes in, audio comes out, and the argument Phonic makes for the architecture is that a transcript is a lossy copy of a conversation — tone, hesitation and overlap get thrown away before the language model ever sees them. That argument has been made by every speech-to-speech vendor since 2024; what Phonic is actually selling is a scheduling trick on top of it.

That trick is tool calling during the caller's turn. Instead of waiting for the caller to stop before it starts a lookup, the model fires tool calls while speech is still arriving, so results are ready by the time the caller yields. In the demo Phonic published, the model called all five tools mid-sentence and answered 1.1 seconds after the caller stopped; it puts OpenAI's GPT-Live-1 at 15.7 seconds on the same script. That comparison is a vendor demonstration on a vendor-selected task, not a measured benchmark, and it is worth reading as an illustration of the design rather than as a head-to-head result.

The published API surface is more concrete than most launch posts:

• Model IDs — phonic-1, with phonic-0.5 still accepted as the previous generation. Omitting the field takes the current default.

• Intelligence — an intelligence_level enum, standard by default, high when you want the more capable model for complex reasoning. The documentation describes standard as the setting optimised for lower latency, which means the configuration every benchmark has scored so far is not the one a latency-sensitive deployment gets by default.

• Languages — 51, listed with ISO codes down to Filipino (fil) and Cantonese (yue), with mid-conversation switching that keeps the same speaker identity.

• Latency — Phonic claims sub-600 ms speech-in to speech-out. Artificial Analysis measured 0.61 s to first audio on its own harness, which is consistent with the claim without being the same measurement.

• Deployment — hosted API, TypeScript and Python SDKs, a CLI and a no-code agent builder, plus self-hosted and VPC options at the enterprise tier, with SOC 2 Type II, a HIPAA BAA, zero data retention and a 99.9% uptime SLA on the page.

The index, row by row

The Speech to Speech Index is a composite of four equally weighted results: Big Bench Audio for speech reasoning, τ-Voice for agentic customer-service behaviour, a human preference arena, and a task-success rate. A model needs all four to be scored at all, which is why StepAudio 3 Realtime appears in the reasoning table without a composite number — its agentic, arena and task-success columns are empty.

Phonic 1's four components, per Artificial Analysis:

• Speech to Speech Index — 87.1, first, ahead of Grok Voice 3.0 at 86.9, GPT-Live-1 (Astra, medium) at 83.2, Grok Voice 2.0 at 81.3 and GPT-Live-1 (Sol, low) at 80.1.

• τ-Voice — 79.9%, first, 5.4 points clear of GPT-Live-1 (Astra, medium) at 74.5%.

• Big Bench Audio — 98.9%, third, behind StepAudio 3 Realtime at 99.7% and Qwen Audio 3.0 Realtime Plus at 99.2%.

• Speech Agent Arena — eighth of 30 on preference at Elo 1012, fourth on task success at 92.3%.

Read those four rows in order and the composite stops looking like a tie. Phonic 1 wins the agentic benchmark outright and places third on reasoning, but it is eighth on which voice humans actually prefer and fourth on raw task success — a model that is measurably the better tool-runner and not the crowd's favourite conversationalist. That is the same structural story we found in Grok Voice 3.0's 86.9 two weeks ago, from the opposite side: xAI held the arena and lost the agentic column, and Phonic holds the agentic column while giving up the arena.

Two clocks, two winners

Latency is where the coverage gets sloppy, because two respectable instruments disagree about who is fastest and neither of them is wrong.

Artificial Analysis measures time to first audio — how long from the end of the caller's turn until audio begins. On that clock Phonic 1 is the fastest of the index's top five:

• Time to first audio — Phonic 1 0.61 s; Grok Voice 3.0 0.81 s; GPT-Live-1 (Sol, low) 1.24 s; GPT-Live-1 (Astra, medium) 1.34 s; Qwen Audio 3.0 Realtime Plus 1.54 s; StepAudio 3 Realtime 8.83 s.

Cekura, a voice-agent evaluation firm that runs models inside a real phone harness over WebRTC, measures median response time across whole calls. On that clock Grok Voice 3.0 is the quickest in the field at 1.34 s and Phonic 1 is 1.63 s — third-fastest, a hundredth behind Grok Voice Think Fast 2.0 at 1.62 s, and not first. Both numbers describe something real. Time to first audio is how fast the model commits to speaking; median response time is how long a caller actually waits for the reply across 82 scenarios, including the tool-heavy ones where Phonic's approach buys back exactly the seconds a first-audio measurement cannot see.

If you are writing a procurement memo, the sentence to avoid is "the fastest speech-to-speech model", full stop. The defensible version is "fastest to first audio on Artificial Analysis; third-fastest median response on Cekura's live-call harness".

The live-call board asks a harder question

Cekura's speech-to-speech leaderboard is a substantially harder test than any of the four index components: 82 scenarios — 59 clinic booking calls and 23 Medicare intake calls — each run three times inside the same open-source Pipecat agent with the same prompt and tools, scored pass-cubed, meaning a scenario counts only if all three runs succeed with every field recorded correctly.

As of its October 22, 2026 update, with Phonic 1 configured at intelligence high under the phonic_v1 model ID, that board also puts Phonic 1 first:

• Reliability, pass-cubed — Phonic 1 90.2%; Grok Voice 3.0 81.7%; GPT Realtime 2.1 79.3%; GPT-Live 1 with GPT-6 Sol 75.6%; Azure Realtime 70.7%.

• Data accuracy — Phonic 1 97.5%; GPT Realtime 2.1 92.0%; Grok Voice 3.0 87.3%.

• Calls where the agent went silent — Phonic 1 0.0%; Grok Voice 3.0 0.0%; GPT Realtime 2.1 3.3%.

• Medicare-call success — Phonic 1 88.4%; GPT Realtime 2.1 81.2%; Grok Voice 3.0 58.0%; GPT-Live 1 with Sol 49.3%.

The Medicare split is the most interesting line in the whole launch. Long, form-filling intake calls are where end-to-end models are supposed to collapse, and Phonic 1 holds 88.4% where Grok Voice 3.0 halves to 58.0%. Cekura's own write-up still logs a failure mode worth knowing: the agent hangs up without acknowledging a caller who says they will call back later, and it occasionally writes a Medicare handoff note or routing detail wrong. At 90.2% pass-cubed, 74 of those 82 scenarios clear every run and eight do not — that is a strong result for a model a day old, not a finished one.

Two labelling notes before anyone quotes these figures. Cekura's board carries a last-updated date and has revised between runs, so a screenshot of it is not a citation; quote the live number with the date attached. And Cekura records a cost of $0.060 per minute for Phonic 1 against a published Starter rate of $0.070 — a measured figure from one harness, not the price list, and not something to subtract from anything.

The vendor benchmarks, kept in their own box

Phonic's launch post carries three results of its own, and they are worth knowing precisely because they are not the numbers above. On Sierra's τ³-Voice multilingual full-duplex variant the company reports Phonic-1 at 83.2 against GPT-Live-1 at 79.9 and Gemini 3.8 Live at 68.5. Note the collision: 79.9 is Phonic's score on Artificial Analysis's τ-Voice and also GPT-Live-1's score on Phonic's own τ³-Voice run. Same digits, different rows, different benchmark variants — never move a number from one column to the other.

The other two are blind human tests run by Podonos, a third-party listener-testing service. Phonic-1 was preferred over GPT-Live-1 by 61.5% to 32.0%, over Grok Voice 3 by 66.5% to 26.5%, and over Eleven v4 Turbo by 78.0% to 14.0%. The result Phonic itself leads with is the one it loses: human voice actors were preferred over Phonic-1 by 57.0% to 32.0%. All four are vendor-commissioned and vendor-reported; none has been independently replicated.

What a month of it costs

Pricing is one number per minute with no per-token charges and no separate LLM pass-through, which is unusual in a category where the "voice model" is often a thin wrapper over three billed services. The tiers:

• Starter — $0.07 per minute, no minimum, 10 concurrent calls, 10 agents, 3 workspace seats.

• Pro — $0.065 per minute against a $200 monthly minimum, which the pricing page equates to roughly 3,100 minutes, with 50 concurrent calls, 50 agents and 30 seats.

• Enterprise — custom, with volume discounts, VPC or on-prem deployment, DPAs, SLAs and HIPAA BAAs.

Sign-ups before November 30 get $250 in credits. Run the arithmetic on a 10,000-minute month and the list prices of the nearest rivals sort themselves out quickly:

• 10,000 minutes — Phonic 1 Starter $700; Grok Voice 3.0 $800 at $0.08/minute; GPT-Live-1 (Sol, low) $750 at $0.075; GPT-Live-1 (Astra, medium) $970 at $0.097.

• Same 10,000 minutes on Pro — $650, but only above roughly 3,100 minutes does the $200 minimum stop being the binding constraint.

The number that does not fit on a price sheet is concurrency. Ten simultaneous calls on Starter is a real ceiling for anything resembling a support queue, and it is the reason the jump to Pro is worth $50 a month at volume: the tier change buys 50 concurrent calls far more than it buys the half-cent. Phonic's own engineering blog argues concurrency limits are the hidden tax in voice AI, which is an honest framing from the party that sets them.

Where it runs, and what routing can and cannot do here

Phonic 1 is not in OrcaRouter's catalogue, and this piece will not pretend otherwise: today you reach it through Phonic's own API, its SDKs, the CLI or the no-code builder, and nowhere else. That is also the honest structural weakness of buying an end-to-end voice model over a cascade — a single vendor, a single endpoint, and no second provider to fail over to when the websocket drops mid-call.

What routing does cover is the alternative. A transcribe-think-speak pipeline is the baseline Cekura's own board keeps beating end-to-end models on, and its middle stage is ordinary text inference — which is exactly the layer where one API for 200+ models with 0% markup applies. The transcription leg and the reasoning leg can sit behind a single key, and the routing DSL can send the cheap turn-taking calls and the expensive policy-reasoning calls to different models on the same call, without a second contract for each vendor in the stack. Provider list price is passed through unchanged, so a vendor price cut reaches that leg the same day it is announced rather than at the next contract renewal.

The trade-off is worth stating plainly, because the scoreboards above do not price it. End-to-end wins on turn-taking, interruption and mid-utterance tool calls; a cascade wins on the ability to swap any one component — and on the fact that Cekura's cascade baseline still outranks every realtime model it tested on pass-cubed reliability. If that sounds like a reason to stay on a cascade, it partly is. The reason to move is the Medicare column: 88.4% against 58.0% is the kind of gap that decides whether a long intake call completes at all.

What nobody has measured yet

Before anyone rewrites a production prompt around an 87.1, three gaps are worth naming.

• The measured configuration is not the default. Every score here — Artificial Analysis's 87.1, Cekura's 90.2% — was produced at the high intelligence setting. The API defaults to standard, documented as the lower-latency mode. Nobody has published the standard-setting numbers, so the latency figures above do not describe what an unconfigured integration experiences.

• Noisy-channel behaviour is untested by the index. Cekura reports Grok Voice 3.0 passing every call with background noise and audio glitches; it says nothing comparable about Phonic 1, and Big Bench Audio runs on clean audio.

• Long-session stability is a demo, not a measurement. Phonic's 44-minute call with 123 tool calls in English and Portuguese is the most concrete long-horizon evidence published, and it is a single call the vendor chose to show.

The defensible summary after one day: Phonic 1 is the strongest agentic speech-to-speech model on the two independent instruments that have scored it, it is not the fastest one depending on which clock you read, and its $0.07/minute makes the agentic lead cheap enough to test rather than argue about.