AssemblyAI-TTS-1 vs Cartesia Sonic 3.6: the Challenger Measured the Incumbent at 143 ms

The most useful sentence in AssemblyAI's launch post is not about its own model. It is a line in a comparison table showing Cartesia Sonic 3.6 taking 143 milliseconds to return a first byte while AssemblyAI-TTS-1 returned one in 47 — a test AssemblyAI ran, on its own hardware, and published about a competitor that has spent two years marketing itself as the fastest TTS on the market. Sonic 3.6's own vendor figure is under 90 ms time to first audio, which is not the same measurement, on not the same network, and the two numbers have been quoted against each other ever since. That collision is the honest centre of this matchup: one model is the quality and language leader at $49 per million characters, the other is the price and throughput challenger at $15, and neither has a latency number that a neutral party has confirmed.

Everything below is dated. Artificial Analysis figures were read on 30 October 2026 and the boards drift by tens of Elo within weeks; Sonic 3.6 has itself moved from first to fourth since August. Vendor claims are labelled as vendor claims. The two independent latency data points that exist — AssemblyAI's lab run and Cekura's streaming measurement — measure different things and are presented as such rather than averaged into a single pretend fact.

Where they actually sit, as of today

Cartesia Sonic 3.6 launched in mid-August 2026 and took first place on both Artificial Analysis speech arenas within days — a 1,282 on Provider Voice and a 1,123 on Controlled Voice, both read at the time. Two and a half months of new entries later it holds 1,296 and fourth on Provider Voice, and 1,147 and sixth on Controlled Voice. Its score went up 14 points and its rank went down three places. That is what a fast-moving board does, and it is why the rank column in any launch write-up ages badly.

• Provider Voice — Cartesia Sonic 3.6 1,296, 4th vs AssemblyAI-TTS-1 1,227, 14th

• Controlled Voice (eight shared cloned voices) — Cartesia Sonic 3.6 1,147, 6th vs AssemblyAI-TTS-1 1,108, 10th

• Price per million characters — $49 vs $15

• Languages — 44 vs 6 (English, Spanish, French, German, Italian, Portuguese)

• Throughput (Artificial Analysis) — 148.6 characters per second vs 162

• Pronunciation robustness (Artificial Analysis) — AssemblyAI 81.7%, 16th vs Sonic's lower placement on third-party hard-string tests

• Cloned-voice billing — Cartesia charges 1.5x for cloned voices vs AssemblyAI's flat rate including cloning

The two gaps are not the same size. Sixty-nine Elo on Provider Voice is a real, visible quality difference — roughly the distance between adjacent models in the middle of the board, and enough that a blind listener has a preference. Thirty-seven Elo on Controlled Voice is smaller and lands in the range where a single evaluation round can move a model several places. In plain terms: Sonic 3.6 sounds better with its own voices, and the two are close enough with a cloned voice that deployment details decide it. A 3.3x price gap is not close by any measure.

The latency argument, told three ways

Three measurements exist and none of them settles the question, so it is worth being precise about what each one shows.

The first is AssemblyAI's own lab test. Thirty prompts, three rounds, four stock voices per model, one request at a time, single client. Median time to first byte: AssemblyAI 47 ms, ElevenLabs Flash v2.5 119 ms, Cartesia Sonic 3.6 143 ms, ElevenLabs v4 612 ms. This is a clean, well-described test with an obvious limitation — sequential requests measure the model, not the service. It cannot show queueing, concurrency contention or cross-region network cost. It also has an incentive problem: the vendor publishing the table is the vendor winning it.

The second is Cartesia's own claim: sub-90 ms time to first audio, which the company describes as model-only and excluding network. Note the unit change. AssemblyAI's 47 ms is a first byte; Sonic's 90 ms is first audio; those are different moments in a streaming pipeline, and a first byte can arrive well before the first synthesised frame. Comparing them directly is the most common error in coverage of this launch.

The third is Cekura's independent streaming measurement, which put AssemblyAI-TTS-1 at 159 ms median across agent-style sentences — eighth of nineteen systems, not first. That is twice the vendor's production figure and roughly in line with what a real network path does to a clean lab number. No equivalent independent figure for Sonic 3.6 was available in the same run, which means the honest statement is not "AssemblyAI is faster" but "AssemblyAI is faster in a test it designed, and neither model has been raced by a neutral lab on the same day."

Throughput is the cleaner comparison because Artificial Analysis measures it the same way for both. At 162 characters per second against 148.6, AssemblyAI generates audio about 9% faster — a much smaller advantage than the first-byte numbers imply, and one that is irrelevant for a live agent where both models are already several times faster than real time. What matters in production is tail latency and reliability under concurrency, and neither vendor has published a p99 figure.

What the price difference includes

Cartesia's $49 is not a flat all-in rate in the way AssemblyAI's $15 is. Sonic 3.6 bills cloned voices at 1.5 times the standard rate, which pushes an agent running custom voices to roughly $73.50 per million characters — about 4.9 times AssemblyAI's rate. AssemblyAI charges $15 for library voices and clones alike, includes word timestamps, US or EU processing and unlimited concurrency, with no tier purchase and no annual commitment. Cartesia sells credits through plans at the free, $5, $49 and $299 monthly marks, and its free tier prohibits commercial use.

Worked through on a script rather than a rate card: a million characters is about 18 hours of finished speech at a normal speaking pace, which makes AssemblyAI roughly 81 cents per finished hour against roughly $2.70 for Sonic 3.6 at list, or about $4.05 once cloned voices are billed at the 1.5x rate. For a hundred-thousand-hour-a-month operation that is the difference between an $81,000 line item and a $270,000 one. Nothing about voice quality justifies a 3.3x multiplier on its own; the case for Cartesia has to be made on coverage, controls and reliability, not on Elo.

Languages and controls, where Cartesia still wins comfortably

Forty-four localised languages against six is the single largest functional gap in this comparison, and it is not a gap that a price cut closes. AssemblyAI ships English, Spanish, French, German, Italian and Portuguese with multiple accents and has promised more without a date. If your agent serves European markets outside those six — Polish, Dutch, Swedish, Greek — Sonic 3.6 is the only option of the two, and the price difference stops mattering.

The controls gap runs the same direction. AssemblyAI-TTS-1 has no SSML, no speed parameter, no pronunciation dictionary and no emotion tags at launch; the vendor's own launch page describes emotion as native to the model while its FAQ confirms dedicated controls are unavailable. Cartesia's product is built around steerable speech — speed, emotion and pronunciation handling are part of the offer. For scripted narration, dubbing or any workload where pacing has to be directed, AssemblyAI currently asks you to rewrite the text to get the delivery you want, which is work you cannot bill to anyone.

The one place AssemblyAI leads on capability is exactly the place Cartesia is weakest: hard strings. Artificial Analysis puts AssemblyAI's pronunciation robustness at 81.7%, sixteenth on that board, and AssemblyAI's own eval puts entity accuracy at 93.7% on phone numbers, URLs and codes against 82.5% for Sonic 3.6 — a vendor comparison, run by the vendor, and an 11-point spread that is too large to dismiss as methodology noise. One independent test (Gradium's) put Sonic's hard-case score at 75.1%. For IVR, readbacks and confirmation codes, that is the number that decides the deployment.

Who should take which

Choose AssemblyAI-TTS-1 when the workload is high-volume, English-plus-five-languages, and heavy on structured readback — order numbers, addresses, codes, dosages — where the entity accuracy advantage and the flat $15 rate compound. Choose Cartesia Sonic 3.6 when you need more than six languages, when delivery has to be directed rather than rewritten, when you are cloning voices and can absorb the 1.5x billing, or when the naturalness gap on Provider Voice is the thing your users hear. The 37-point controlled-arena difference is small enough that a well-built agent on either model will sound fine; the 69-point provider difference is the one a listener notices.

There is a third option that neither vendor advertises, and it is the one most teams land on: run both and let traffic decide. Two TTS legs behind one routing rule lets you send readback-heavy utterances to the model that reads codes correctly and narration to the model that sounds better, without maintaining two integration paths. That is the case a routing layer makes for itself — more than 200 models behind one API at 0% markup with list prices passed through, so a cut on either side lands the same day, and the routing DSL to split traffic by workload rather than by guesswork. Neither AssemblyAI-TTS-1 nor Cartesia Sonic 3.6 is in the OrcaRouter catalogue today, so this is a shape for the comparison rather than a claim about hosting them; the TTS models we do route — OpenAI's tts-1, tts-1-hd and gpt-4o-mini-tts, and Google's Gemini 3.1 Flash TTS Preview — cover the second leg while the two leaders are evaluated on your own traffic.

What would actually settle it

A neutral lab running both models on the same hardware, the same network path and the same load profile would end the latency argument in an afternoon, and nobody has done it. Until then the defensible position is that AssemblyAI is faster in its own test and possibly faster in reality, Cartesia is the better-sounding model with far broader coverage, and the price gap is large enough that Sonic 3.6's quality advantage has to be worth 3.3x per character — or, on cloned voices, 4.9x — to the buyer paying for it.

Two other developments would move the answer more than any benchmark. If AssemblyAI ships SSML and pronunciation dictionaries, the six-language limit becomes the only remaining functional objection and the price case gets very hard to argue with. If Cartesia cuts its rate toward the market's new floor, or extends its cloned-voice billing, the coverage advantage becomes affordable. Watch those two moves rather than the next Elo reading — on a board where a model can gain 14 points and lose three places in ten weeks, the rank is noise and the price is signal.