Rime Prism vs Cartesia Sonic 3.6: Stop Comparing 90 ms to 570 ms

Cartesia Sonic 3.6 answers in under 90 milliseconds. Rime Prism answers in about 570. If that sentence is the whole of your comparison, you have already reached a conclusion that is off by a factor of two in the wrong direction — because the two numbers do not start at the same moment. Sonic 3.6's 90 ms is the time from text arriving at the synthesiser to the first audio sample. Rime Prism's 570 ms is the time from a caller finishing their sentence to the agent making a sound. Between those two events sit a speech-recognition pass and a full language-model turn, neither of which Sonic 3.6 performs. Put them on the same clock and the gap narrows to something a caller can barely distinguish, and then the decision becomes about price, language coverage and whether anyone has to approve what gets said.

Both models are current and both are shipping. Cartesia Sonic 3.6 went GA on August 27, 2026 as a streaming text-to-speech model that took the top of both Artificial Analysis speech arenas at launch — and has since been passed as newer voices joined the board. Rime Prism launched October 28, 2026 as a speech-to-speech model at $0.080 per minute of call audio that emits its reply as text before speaking it. One is a component, one is a system, and the interesting comparison is what each costs when you assemble a working voice agent around it.

What Sonic 3.6 is, precisely

It is a text-to-speech model, and a good one by the available evidence. The streams matter more than the marketing: WebSocket, Server-Sent Events and raw bytes are all supported, which is what makes a TTS model usable inside a live call rather than only in a batch job.

• Coverage — 44 languages across 61 locales, including Hindi and Hinglish.

• Voices — 10-second instant cloning from a reference clip; no published count of stock voices for this release.

• Control — more than 60 emotion and delivery tags, English only; speed from 0.6x to 1.5x.

• Audio — 8 kHz up to 48 kHz output in WAV, MP3 or PCM, with a 3,000-token context window.

• Latency — sub-90 ms vendor-stated time to first audio at roughly 132 characters per second, streaming.

• Arena — Artificial Analysis Provider Voice Elo around 1,282 and Controlled Voice around 1,123 at launch, top of both tables then, since moved.

• Openness — no weights. Sonic 3.6 is a hosted model, like Prism.

Cartesia also retired several older models on October 20, 2026 — sonic-turbo, sonic-2 and the sonic-3-2025-10-27 snapshot — and Sonic 3.5 remains listed as stable alongside 3.6. If you are migrating, check which generation your integration actually points at, because a fallback to a retired model is a silent failure.

The 90 ms vs 570 ms arithmetic, done properly

Here is the part every ranking page skips. A Sonic 3.6 cascade has three clocks running in series:

• Speech recognition of the caller's utterance — typically a few hundred milliseconds for a streaming recogniser to finalise a sentence.

• The language model's turn — the variable term, anywhere from 200 ms for a short acknowledgement to well over a second if the agent reasons before replying.

• Sonic 3.6's synthesis hop — under 90 ms at the vendor's tested rate with network excluded; independent end-to-end measurements of a live Cartesia call path land several times higher once the other two hops are counted.

Add the middle three and a well-tuned cascade lands somewhere between 800 ms and 1.5 seconds from end of caller speech to first agent audio. Prism's 0.57 seconds covers all of it. So on the complete path, Prism is not slower than Sonic 3.6 — it is faster than the pipeline Sonic 3.6 lives inside, by roughly a factor of two at the good end.

That flips only in one architecture: if your language-model turn is very short and your recogniser is very fast, a Sonic cascade can reach first audio in around half a second too, and then the 90 ms hop is genuinely the reason. Which is to say the 90 ms is a real engineering achievement and it is doing work — just not the work the comparison table implies.

The honest framing is the one Cartesia's own numbers support. Sub-90 ms is the fastest synthesis hop on the market, and it is measured with network excluded. Prism's 570 ms is a vendor figure, measured on the τ-Voice harness by Rime itself rather than by a neutral lab. Neither company has published a head-to-head, and nobody has run both on the same call path. Anything that presents the two as a like-for-like latency race is inventing a result.

Price per minute, which is where the real gap lives

Cartesia bills by character: roughly one credit per character, which Artificial Analysis normalises to about $49 per million characters. Rime's own documentation converts characters to minutes at roughly 1,000 per minute of speech, which puts Sonic 3.6 near $0.049 per minute of generated audio. Faster speaking rates — the 132 characters per second Cartesia benchmarks at — push that down toward $0.034. Call it three to five cents a minute, depending on how quickly your agent talks.

Prism is $0.080 per minute of call audio on Flex and $0.064 on Scale. But Prism's minute is not the same minute. It is billed on call audio, not on agent speech, so a call where the caller talks half the time costs Prism the same as one where the agent never stops. That distinction alone can shift the effective comparison by a third in either direction, and it is the first thing to model when you price a trial.

Now assemble the cascade. Sonic 3.6 at four cents, recognition at half a cent to a cent, and a language model that costs a few cents a minute at conversational token volumes — you are at roughly eight to ten cents per agent-spoken minute, before any infrastructure. Prism at eight cents is not a premium product in that frame. It is the same price for one bill, one contract and one latency budget instead of three.

Cartesia's plan structure reinforces this. Free with 20,000 credits and 2 concurrent; Pro at $5 for 100,000 credits and 3 concurrent; Startup at $49 for 1.25 million credits and 5 concurrent; Scale at $299 for 8 million credits and 15 concurrent; Enterprise custom with DPA, BAA and regional deployment. Concurrency is capped by tier, which for a contact center is the number that actually binds — 15 concurrent streams on Scale is a busy mid-size queue, not a large one.

Language coverage versus control: the trade that decides it

Cartesia wins coverage by a wide and measurable margin. Forty-four languages across 61 locales is a published, testable list. Rime does not publish Prism's language count at all, though its separate text-to-speech products publish theirs — but Prism-specific figures are absent from the product page. If you need to launch in fifteen markets next quarter, that absence is disqualifying on its own, and Sonic 3.6 is the answer regardless of anything else in this article.

Prism wins control, and by a margin that is structural rather than incremental. Sonic 3.6 receives final text. It can perform that text with more than 60 emotion tags and it can say it at 1.5x speed, but it cannot refuse it. Prism produces the text itself, one step before speech, and Rime says the application can rewrite, block, tag or correct pronunciation at that point while the call continues. A regulated disclosure, a claim that must not be made, a competing brand name that must never be spoken — those live at the checkpoint, not in the text you hand a synthesiser.

You can build the same control in a Sonic cascade. Put a policy check between the language model and the synthesiser, and you have the checkpoint. What you do not have is a model that reasons over audio in the same pass, and that is the second half of Prism's claim: tone, hesitation and background noise inform the reply rather than being flattened into a transcript first. Whether that is worth eight cents versus four is a question about your calls, not about the models.

Both of them stream, neither of them is on our catalogue

Prism is sold direct by Rime. Sonic 3.6 is sold direct by Cartesia and through several third-party platforms. Neither is routed by OrcaRouter today, and nothing in this piece should be read as claiming otherwise.

What we do route is the layer both of them depend on. In a Sonic cascade, the language model in the middle is a choice, and that choice is where the quality and the cost of the call actually live. One key on OrcaRouter covers 200-plus models at 0% markup over provider list price, with automatic failover between providers, so a conversational model that stalls does not become dead air — the request reroutes instead. The routing DSL lets you send the acknowledgement turns to a small cheap model and the lookup turns to a larger one, inside one endpoint, which is the same cost discipline a cascade needs and the reason a cascade is still competitive on price.

And if you want a hosted synthesis hop without a second vendor relationship, openai/tts-1-hd and google/gemini-3.1-flash-tts-preview are both on the catalogue at list price. They are not Sonic 3.6. They are the other two ends of the same pipeline, on the same key.

The verdict, split by architecture

If you are building in more than a handful of languages, Cartesia Sonic 3.6 is the correct synthesiser and the comparison is over. Forty-four languages is a fact; an unpublished list is not a promise.

If you are building a single-market contact center where the agent must not say certain things, Rime Prism is the only one of the two with a place to stand between the decision and the utterance, and at eight cents a minute it is roughly the same money as the cascade you would otherwise assemble — with one bill instead of three.

If you are somewhere in between, the tiebreaker is latency under load rather than latency on a benchmark chart. Cartesia's sub-90 ms hop is measured with network excluded, and a deployment far from the region will see considerably more. Prism's 570 ms is measured on a harness that has no network at all. Run both on your actual call path, from your actual region, before you pick. The published numbers cannot settle this, and anyone who says they can is comparing two different stopwatches.

One caveat on the arena numbers, because they will get quoted: Artificial Analysis Elo for Sonic 3.6 measures how a voice sounds to human raters. Prism's 56.6 on τ-Voice measures how often an agent completes a task on 278 simulated service calls. They are not the same instrument, they do not share a scale, and subtracting one from the other produces a number that means nothing. Sonic 3.6 has no τ-Voice score because it is not an agent. Prism has no arena Elo because it is not a voice. That asymmetry is the entire comparison in miniature.