Phonic 1 vs Cartesia Sonic 3.6: A $0.07 Minute and a $49 Million-Character Bill Are Not the Same Purchase

Put Phonic 1 and Cartesia Sonic 3.6 on one page and the temptation is to line up the price columns and call it a day. Two vendor numbers make that easy: Phonic charges $0.07 for a minute of conversation, Cartesia normalises to about $49 per million characters. Both sound like rounding errors next to what you were paying in 2025. But they are billing for different objects. Cartesia Sonic 3.6 is a streaming text-to-speech engine — you hand it words, it hands you a voice, and everything upstream of those words is your problem. Phonic 1 is a speech-to-speech model that hears a caller, decides what to do, calls tools mid-sentence and answers out loud, all inside that minute. One is a component on your invoice. The other is the whole invoice. Comparing them on price alone is comparing a taxi fare to a car.

That does not make the matchup useless. It makes it narrower and more honest than the search result suggests: these two compete for the same line item in a voice-agent budget, and a team building a phone agent has to decide whether the voice layer is a service it rents whole or a stack it assembles from parts. Cartesia will sell you either arrangement — Sonic 3.6 for the rendering half, Cartesia Line for the assembled half — and that second product, not the TTS model, is the one actually sitting across from Phonic 1.

What each one actually is

Cartesia Sonic 3.6 shipped in late August 2026, roughly three months after Sonic 3.5, and it is the company's highest-quality renderer: 44 languages across 61 locales, voice cloning from about ten seconds of audio, credit-based billing where one credit approximates one character. It has no reasoning layer, no tool layer and no notion of a conversation. You call it with text and a voice ID, you get audio back in a stream, and the turn-taking logic — when to listen, when to interrupt, what to do when the caller says "actually, no" — belongs to code you write.

Phonic 1 launched in late October 2026 alongside a $27.5M Series A led by First Round, and its pitch is the opposite shape: a single model that owns the whole turn. The vendor's own feature list claims tone and hesitation are heard, tool calls fire while the caller is still speaking, and the model will sometimes choose not to respond at all — a behaviour you cannot get from a TTS endpoint because a TTS endpoint has no opinion about whether speaking is the right move. It bills one price per conversation minute and states explicitly that there is no separate LLM pass-through and no per-token charge.

One honest wrinkle before we go further. Phonic's launch coverage described it as end-to-end, avoiding separate speech-recognition, LLM and speech-synthesis vendors. The vendor's own site describes supervisor models for reasoning and tool calling, and its benchmark configuration is labelled Phonic 1 (High), implying a selectable intelligence level. Both can be true — a single vendor can own the whole turn while still running an internal reasoning layer — but "end-to-end" is marketing shorthand for "one contract", not proof that a single network does everything in one forward pass. If your architecture review depends on that distinction, ask the vendor directly; the public material does not settle it.

The two clocks measure different things

Latency is where this comparison misleads most often, because both vendors quote a small millisecond number and neither number is the one your caller experiences.

• Phonic 1 — 0.61s average time to first audio on Artificial Analysis's measurement, fastest among the top five speech-to-speech models tested; the vendor advertises sub-600ms from speech-in to speech-out.

• Cartesia Sonic 3.6 — sub-90ms time to first audio, vendor-stated and explicitly excluding network; a third-party benchmark measured 120ms p50 from US East against the vendor directly, and a Sydney-based real-call test reported 448ms median and 604ms p90 end to end.

• What the gap means — Phonic's 0.61s starts at the moment audio arrives and includes the decision about what to say. Cartesia's 90ms starts once the text already exists and only covers turning it into sound. Add your own speech recognition and your own model latency in front of Sonic and the two figures converge; that missing arithmetic is why the raw numbers look like a seven-to-one win and are not one.

Artificial Analysis also reports Phonic 1 at 0.61s against 1.34s for GPT-Live-1 and 1.35s for Gemini 3.8 Live Extended Thinking, which is the comparison the number is actually good for: model-against-model, same instrument. Sonic 3.6 does not appear on that board at all, because it is not a conversational model.

The bill, built from the same job

Take a ten-minute inbound support call. On the assembled path with Cartesia Sonic 3.6 you pay for at least three things: recognition of the caller's audio, a model to decide the reply, and the rendering of that reply. Sonic's share is the smallest and most predictable of the three — around $0.28 to $0.49 per minute of generated audio depending on plan, and noticeably less than that in practice because an agent does not speak for the full ten minutes. The unpredictable share is the model in the middle, and it is the one that varies by an order of magnitude depending on whether you route a small model for the easy 80% of turns or send everything to a frontier model.

That middle leg is where routing earns its place rather than being a bolt-on. On OrcaRouter the reasoning model behind your agent sits on one key across 200+ models at 0% markup over provider list price, so a price cut from any vendor is live the same day rather than waiting on a contract amendment — and a routing rule can keep cheap turns cheap while escalating only the hard ones. Automatic failover covers the case that actually breaks voice agents in production: the model provider blips mid-call and the caller hears silence.

On the Phonic 1 path the arithmetic collapses into one line. Ten minutes is $0.70 at the $0.07 starter rate, or $0.65 at the Pro rate, which carries a $200 monthly minimum. That is the whole cost of the model layer — no separate tokens, no separate reasoning bill — and it is the strongest argument in Phonic's column, because a flat per-minute rate is a number a finance team can forecast without a usage model.

Read those two totals carefully before declaring a winner. The assembled path's floor is lower — a cheap middle model plus Sonic's rendering can beat $0.07 a minute comfortably — and its ceiling is much higher, because a frontier model at every turn blows past it. Phonic 1's flat rate is a bet that the vendor's economics hold as your mix of easy and hard calls changes. That is a genuinely useful thing to buy, and it is also a thing you are buying on trust.

Where Cartesia Sonic 3.6 wins without an argument

If the question is "which of these sounds better", Cartesia is not close, and no amount of agentic benchmark scoring changes it. Sonic 3.6 took the top of both Artificial Analysis speech arenas on release — a Provider Voice Elo around 1,282–1,283 and a Controlled Voice score near 1,123 — which are preference measurements over rendered audio, exactly the thing a TTS engine is for.

The control surface is also deeper. Voice cloning from ten seconds, 44 languages across 61 locales with locale-level selection, five published plan tiers from a free 20K-credit sandbox to an $299 Scale plan with fifteen concurrent requests, and cloned voices billed at 1.5x. If your product is narration, dubbing, IVR prompts, or an assistant whose replies are scripted and whose voice is the brand, Sonic 3.6 is a better tool than Phonic 1 and the comparison should stop there.

There is also a cost shape worth naming: Cartesia's credit model means you pay for characters you render, so a call where the agent says four words costs four words. A per-minute model charges for the whole minute regardless. On short, high-volume interactions — appointment confirmations, delivery notifications, a two-turn verification — the character model is structurally cheaper.

Where Phonic 1 wins without an argument

Tool calls that start before the caller stops talking are the one capability in this matchup that Cartesia Sonic 3.6 simply does not have, because tool calling is not a property a renderer can possess. In a booking flow where the caller says "I need Thursday, no wait, Friday afternoon" the assembled path has to wait for end-of-turn detection, run recognition, run the model, run the tool, then render — four serial handoffs against Phonic's claimed one. Phonic's own demo comparison, a response 1.1 seconds after the caller stopped against 15.7 seconds for OpenAI's GPT-Live, is a vendor demonstration and should be read as one. But the structural point survives the caveat: fewer handoffs is fewer places to lose a second.

• Languages — Phonic 1 supports 50+ with mid-conversation switching that keeps the same voice; Cartesia Sonic 3.6 covers 44 languages across 61 locales for rendering only.

• Billing — Phonic 1 is one price per conversation minute with no LLM pass-through; Cartesia Sonic 3.6 bills per character/credit, and the rest of the stack bills separately.

• Agentic capability — Phonic 1 scores 79.9% on the agentic component of Artificial Analysis's speech-to-speech index, the highest recorded there; Sonic 3.6 has no agentic score because it has no agent.

• Voice quality — Cartesia Sonic 3.6 leads both Artificial Analysis speech arenas on rendered audio; Phonic 1's arena preference score is 1,012, eighth of thirty.

• Deployment — Phonic 1 offers cloud, single-tenant, VPC peering and self-host, with SOC 2 Type II and HIPAA BAA availability stated; Sonic 3.6 is a hosted API.

The comparison Cartesia would actually want you to run

Cartesia's real answer to Phonic 1 is Cartesia Line, its voice-agent product, which the company prices at about $0.06 per minute and which appears on Artificial Analysis as a provider pipeline rather than a scored model — the line-up behind it draws on Ink, Gemini 2.5 Flash and Sonic together. That tells you two things at once. The headline rate undercuts Phonic 1 by a cent, and the rate is incomplete in exactly the way Phonic's marketing accuses assembled stacks of being incomplete: the Gemini 2.5 Flash leg bills its own tokens on top.

So the honest head-to-head is not $0.07 against $0.06. It is a flat, all-in $0.07 against a $0.06 platform fee plus a model bill you control. For a team that already routes its own models and wants to keep doing so, that structure is an advantage. For a team that wants a number, not a spreadsheet, it is the reason Phonic 1 exists.

Who should act, and who should wait

Buy Cartesia Sonic 3.6 if the voice is the product. Narration, branded assistants, dubbing, anything where a human will judge the audio and nobody will judge the tool calls. The arena lead is real, the cloning is fast, and the plan ladder means you can start for free.

Buy Phonic 1 if the call has to do something — book, verify, refund, escalate — and if a five-figure engineering budget for glue code is worse than a flat per-minute rate. It is also, at the time of writing, roughly a week old on a public API with one independent board behind it and no long-run reliability data; the sensible way to adopt it is beside your current path, not instead of it.

And if you are assembling rather than renting, the decision that actually moves your bill is not Sonic versus Phonic. It is which model answers the easy turns. That leg is the one that runs on OrcaRouter today — one key, 200+ models, list price passed through, failover when a provider drops — while neither Phonic 1 nor Cartesia Sonic 3.6 is on our catalogue, and each must be contracted from its own vendor.

The purchase that looks like a price comparison is really a question about how much of the turn you want to own. Sonic 3.6 lets you own all of it and gives you the best-sounding component to build on. Phonic 1 takes it off your hands and charges a flat rate for the privilege. Neither is wrong. They are wrong only when bought for the other's reason.