AssemblyAI-TTS-1 at $15 a Million Characters: the Voice Agent Price Floor Just Moved

AssemblyAI-TTS-1 went generally available on 27 October 2026 at $15 per million input characters, and the number that matters is not the one in the headline — it is the one you get after dividing. At a normal speaking pace of roughly 15 characters per second, a million characters is about 18 hours of finished audio, which puts the model near 81 cents per hour of speech. Cartesia Sonic 3.6, the incumbent speed leader, bills $49 per million characters, or about $2.70 an hour. ElevenLabs v4 bills $80, or about $4.40 an hour. This is AssemblyAI's first standalone text-to-speech model — the company was a speech-to-text vendor until this week — and it has walked into a market where the leader charges three to five times more for the same unit of work.

The claim deserves an immediate qualifier, because three different parties are supplying three different kinds of number here. The price is AssemblyAI's own published rate card and is checkable on its pricing page. The latency figure is from a lab test AssemblyAI ran against five rival providers, on 30 prompts, three rounds, four stock voices per model, one request at a time. The leaderboard placement is Artificial Analysis's, measured independently. Where this article cites AssemblyAI's own quality evals — entity accuracy, predicted MOS, prosody, clone similarity — those are vendor-reported and nobody outside AssemblyAI has reproduced them. The one genuinely independent latency measurement available so far points somewhere less flattering than the marketing, and it is worth leading with that.

The spec sheet in one screen

• Price — $15 per million input characters, billed on text including spaces and punctuation, all library voices and all languages included, no per-voice or concurrency fee

• Voices — more than 4,000 in the library, plus instant cloning from a short reference sample (the demo used 24 seconds of audio)

• Languages — English, Spanish, French, German, Italian, Portuguese at launch, multiple accents, no date given for the next batch

• Latency — 90 ms time to first audio quoted for production, described by the company as conservative; 47 ms median time to first byte in its own lab comparison

• Throughput — 162 characters per second on Artificial Analysis, ahead of Cartesia Sonic 3.6 at 148.6

• Interfaces — HTTP, WebSocket with a cancel message for barge-in, and an OpenAI-compatible route; LiveKit and Pipecat reference agents ship with it

• Output — PCM, mu-law and A-law at 8 to 48 kHz, 24 kHz native, word-level timestamps returned with the audio

• Residency and limits — US or EU processing at the same price, unlimited concurrency on pay-as-you-go, no annual contract or tier purchase

• Not yet — no SSML, no speed parameter, no pronunciation dictionaries, no emotion tags, no fine-tuned "pro" cloning

Two of those lines are doing more commercial work than the rest. Unlimited concurrency at list price removes the negotiation that most TTS vendors put in front of high-volume voice agents, where per-minute concurrency tiers can outrun the per-character rate. And billing that counts spaces and punctuation is honest but not free: a script of 1,000 words runs roughly 6,000 characters once spaces and commas are included, so the effective price per spoken word is higher than a naive comparison against per-word pricing suggests. AssemblyAI is not hiding that — the pricing page says it plainly — but it is the kind of detail that turns a 3x headline gap into a 2.7x one on a real script.

Why the 47 ms number is carrying so much weight

The launch post leads with speed, and the underlying test is more informative than most vendor comparisons because it names its competitors and its method. Median time to first byte, single client, sequential requests: AssemblyAI-TTS-1 at 47 ms, ElevenLabs Flash v2.5 at 119 ms, Cartesia Sonic 3.6 at 143 ms, Inworld TTS-2 Flash at 244 ms, Inworld TTS-2 at 280 ms, ElevenLabs v4 at 612 ms. If you accept that setup, AssemblyAI is not merely fast — it is roughly three times faster to first byte than Cartesia Sonic 3.6, the model it prices against, and two and a half times faster than ElevenLabs Flash v2.5.

What the setup cannot tell you is what happens under load. A single client issuing one request at a time measures the model, not the service: no queueing, no concurrency contention, no cold-start path, no network hop across regions. The 90 ms production figure is the company's own attempt to account for that, and the gap between 47 and 90 ms is roughly the size of the real-world penalty the lab test cannot see. Anyone who has run a voice agent at volume will recognise this as the standard shape of TTS latency marketing: a clean lab number for the benchmark slide, a softer production number for the pricing conversation, and a third number that only appears once you are on a public network in a region far from the endpoint.

That third number does exist now, and it is worse. Cekura, an independent voice-agent evaluation outfit, measured a median streaming latency of 159 ms for AssemblyAI-TTS-1 across its test set of agent-style sentences — placing it eighth of nineteen systems rather than first. Cekura also scored reading accuracy at 81.5%, essentially level with ElevenLabs v4 at 80.4%, while ranking AssemblyAI first on phone numbers and codes at 91.7% and fourth on naturalness at 4.50. Read the two sources together and a coherent picture forms: AssemblyAI-TTS-1 is genuinely quick to start speaking, unusually good at hard-to-pronounce tokens, and not the fastest thing on the board once a real streaming path is in the loop.

Two Artificial Analysis arenas, two different stories

Artificial Analysis runs two TTS boards and they measure different things. The Provider Voice arena has each model speak with its own native voices; the Controlled Voice arena forces every model onto the same eight cloned voices. That distinction is the most useful thing in this launch, because AssemblyAI-TTS-1 scores better on the controlled board than on its own catalogue.

• Provider Voice — 1,227 Elo, 12th on debut, 1,626 appearances rated. Two days later the live board reads 1,227 at 14th: same score, lower rank, because two models entered above it. Elo is the stable number; rank is a moving target.

• Controlled Voice — 1,110 Elo, 10th on debut, 1,674 appearances. The live board now reads 1,108 at 10th.

• Pronunciation robustness — 81.7% across 2,115 rated clips, 16th on that board.

• Generation speed — 162 characters per second.

A model that ranks higher among cloned voices than among its own 4,000-voice library is telling you something specific: the acoustics are competitive, and the voice design is not yet the differentiator. Breeze TTS 2, the open-weights leader, shows the exact opposite pattern — strong on provider voices, weak on controlled ones — and that comparison is where this gets interesting. AssemblyAI has built a broad, cheap, competent catalogue; it has not yet built voices that win blind listening tests on their own merits.

The rank drift is worth noting for a practical reason. Every aggregator covering this launch is quoting "top 12 in Provider Voice" from Artificial Analysis's announcement post, and by the time you read this the same score may sit at 15th or 16th. That is not AssemblyAI losing quality. It is a leaderboard where a handful of Elo points spans five places, and where new entries arrive most weeks. Judge the Elo, note the date you read it, and do not treat a two-place slide as a regression.

Where the vendor evals look strong, and where they do not

AssemblyAI published a fuller set of numbers than most TTS launches, all of them its own work, all tested on 50 shared cloned voices with generated audio transcribed back through its Universal-3.5 Pro recogniser. The methodology is sensible — round-tripping audio through an ASR model is a reasonable proxy for intelligibility — and it is still a test the vendor designed and scored.

• Entity accuracy on hard targets (phone numbers, URLs, confirmation codes, drug names) — 93.7% for AssemblyAI-TTS-1, 93.6% for ElevenLabs v4, 82.5% for Cartesia Sonic 3.6, 68.0% for ElevenLabs Flash v2.5. A tenth of a point over ElevenLabs v4 is a tie wearing a win's clothing; the 11-point gap over Sonic 3.6 is the real finding.

• Word accuracy — 99.0%, which is close enough to the ceiling that it stops discriminating between good models.

• Predicted MOS — 3.85, and here AssemblyAI loses: Inworld TTS-2 Flash scores 3.88 on the same scale, with Cartesia Sonic 3.6 at 3.82 and both ElevenLabs models below them. The vendor's own table does not put it first.

• TTSDS prosody — 85.5, a figure with no cross-vendor baseline in the post.

• Clone speaker similarity — 0.64, which is low for instant cloning and consistent with the controlled-arena placement rather than the provider one.

The honest summary of that table is narrower than the launch copy: AssemblyAI-TTS-1 is excellent at getting difficult strings right, competitive on intelligibility, mid-pack on naturalness, and weakest on voice identity when you clone. For a voice agent reading account numbers and postcodes, the first line is the one that pays. For a narrated audiobook or a character voice, the last two lines are the ones that matter, and this model is not the pick.

What it cannot do yet, and what that costs you

The feature gaps are ordinary for a first release and unusually consequential for this category. There is no SSML, so no reliable way to force a pause, an emphasis or a pronunciation without rewriting the text. There is no speed parameter, which means pacing has to be handled upstream by generating more or fewer words. There are no custom pronunciation dictionaries, despite the entity accuracy being the headline strength — when the model does mis-say a brand name, you cannot hand it a lexicon, you can only respell the text and hope. Emotion tags and fine-tuned pro cloning are both "coming soon" in the vendor's own words, and the launch page's "native emotion" heading sits awkwardly next to an FAQ that says dedicated emotion controls are unavailable.

Language coverage is the sharper limit. Six languages at launch against 44 for Cartesia Sonic 3.6 and a claimed 50 for Breeze TTS 2 — with the caveat that Breeze's own model card tags only English and Chinese. If your agent needs to answer a German caller and a Portuguese caller in the same week, AssemblyAI covers both. If it needs Polish, Japanese or Arabic, it does not, and no date has been given.

How you actually get it

AssemblyAI-TTS-1 is served from the vendor's own API — HTTP, WebSocket, and an OpenAI-compatible endpoint — with US and EU hosts at the same price, plus a $50 signup credit and a listed starting point of pay-as-you-go only. Beyond that it is available through several third-party platforms. It is not in the OrcaRouter catalogue today, and we are not going to pretend it is; if you want it this week, the vendor's API is the direct route.

The part of this launch that does concern us is the shape of the decision it creates. A voice agent is a three-model pipeline — speech recognition, a language model, speech synthesis — and the TTS leg is now the one with the widest price spread and the newest entrants. That is exactly the leg where you want a second option wired in before you commit, because a brand-new model with a single-client latency test and no independent load testing is the definition of a component that should sit behind a fallback. Routing more than 200 models through one API at 0% markup is what we do, with the vendor's list price passed through so a price cut lands the same day and automatic failover so a bad afternoon at one provider does not take the agent down. The TTS models we route today are OpenAI's tts-1, tts-1-hd and gpt-4o-mini-tts alongside Google's Gemini 3.1 Flash TTS Preview — a workable second leg while AssemblyAI-TTS-1 proves itself under your traffic rather than its own.

What would change the verdict

Three things, in order of how much they would move the decision. First, an independent latency run under real concurrency: Cekura's 159 ms streaming figure is the only outside measurement so far, and it is a single number from a single lab. A second lab reproducing the 47 ms first-byte advantage — or failing to — would settle whether the speed claim survives contact with a queue. Second, the pronunciation dictionary. The entity accuracy is the model's best argument, and it is also the capability most likely to be undone in production by a product name the model has never seen. Shipping dictionaries turns a benchmark win into an operational one. Third, whether the six languages grow on the same schedule as the rest of the roadmap, because the price advantage only compounds if the coverage catches up.

Until those land, the correct reading of this launch is narrow and useful: AssemblyAI-TTS-1 is the cheapest credible TTS API on the market from a vendor with a decade of speech infrastructure behind it, it is fast in a lab and fast enough in the one outside test that exists, it reads hard strings better than anything it was compared against, and it is mid-pack on naturalness and weak on cloned identity. That is a strong buy for high-volume agents in six languages, and a premature one for anything that needs a distinctive voice.