AssemblyAI-TTS-1 vs ElevenLabs Eleven v4: What 5.3× Cheaper Buys, and What It Costs

AssemblyAI-TTS-1 costs $15 per million characters. ElevenLabs Eleven v4 costs $80. That ratio — 5.3× — is the number every headline about the matchup leads with, and it is the least useful number in it, because the model you would actually put behind a phone agent is not Eleven v4. It is Eleven v4 Turbo, at $40 per million characters, or Flash v2.5, also $40. Against those the discount is 2.7×, and against Flash v2.5 — the low-latency sibling ElevenLabs sells for exactly this job — AssemblyAI-TTS-1 is not fighting a legacy incumbent at all. Both shipped in the last five weeks. Both were built for the same customer. One of them is paying for a quality ceiling the other has decided not to compete on.

The price table is doing more work than the benchmark table

AssemblyAI-TTS-1 went generally available on 27 October 2026, with evaluation numbers published a day later. Its rate card is unusually flat: $15 per million input characters, spaces and punctuation included, and that single rate covers every voice in the 4,000-plus library, every launch language, instant voice cloning, word timestamps, and either US or EU processing. There are no per-voice fees, no concurrency tiers, no annual contract, and the vendor says the per-character price is the same whether you run ten concurrent calls or ten thousand.

ElevenLabs' card is tiered by model. Eleven v4, announced 28 September 2026, lists at $0.08 per thousand characters — $80 per million. Eleven v4 Turbo and Flash v2.5 both list at $0.04 per thousand, or $40 per million. A launch promotion cut v4 by 72% until 12 October 2026; that window has closed, so the $80 is now the number you would actually be billed.

Translate both into the unit a voice-agent team budgets in. At a normal speaking rate of about 150 words per minute and roughly six characters per word including spaces, a minute of finished speech is around 900 characters. That is an estimate, not a vendor figure, but it is close enough to see the shape of the bill:

• One hour of continuous speech — about 54,000 characters — costs roughly $0.81 on AssemblyAI-TTS-1, $4.32 on Eleven v4, and $2.16 on Eleven v4 Turbo.

• A hundred thousand three-minute calls — 300,000 minutes, about 270 million characters — costs about $4,050 on AssemblyAI-TTS-1, $21,600 on Eleven v4, and $10,800 on Eleven v4 Turbo.

• The difference between AssemblyAI-TTS-1 and Eleven v4 Turbo on that workload is $6,750 a month for identical call volume. Against Eleven v4 it is $17,550.

That is the real decision, and it is a boring one: you are choosing whether a quality delta is worth $6,750 a month at a hundred thousand calls. Everything else in this matchup is about how big that delta actually is.

What the extra $65 per million characters buys

On the blind-listening board, it buys a measurable amount. Artificial Analysis's Provider Voice Arena, read on 30 October 2026, ranks Eleven v4 Turbo first at 1,336 Elo and Eleven v4 second at 1,334, against AssemblyAI-TTS-1 at 1,227 in fourteenth place. A 107-point gap on that scale is not noise — the intervals do not overlap. On the Controlled Voice Arena, which holds the reference recording constant so the comparison is about cloning fidelity rather than stock voices, AssemblyAI-TTS-1 sits tenth at 1,110.

The rest of the gap is inventory and control surface, and it is larger than the Elo difference suggests:

• Languages — AssemblyAI-TTS-1 launched with six: English, Spanish, French, German, Italian and Portuguese, with multiple accents each. Eleven v4 covers 90-plus languages with native accents rather than an English speaker reading foreign words.

• Voice library — 4,000-plus on AssemblyAI-TTS-1 against 17,500-plus advertised for Eleven v4.

• Cloning — instant cloning only on AssemblyAI-TTS-1, from a sample the vendor demonstrates at 24 seconds. Eleven v4 pairs instant cloning from as little as ten seconds with Professional Voice Cloning, which returned in v4 after being absent from v3. Existing professional clones may need v4-specific retraining.

• Controls — AssemblyAI-TTS-1 ships with no SSML, no speed parameter, no pronunciation dictionary and no emotion tags. Its launch post is explicit that punctuation and written disfluencies are read as written and that pauses are induced by prompting, not markup. Eleven v4 ships stackable inline audio tags — [laughs], [whispers] — and a 10,000-character request limit.

• Output and plumbing — AssemblyAI-TTS-1 streams over WebSocket and HTTP, exposes an OpenAI-compatible endpoint, ships word timestamps, supports a cancel message for barge-in, and publishes LiveKit and Pipecat reference agents. Eleven v4 is sold through ElevenAPI, ElevenAgents and ElevenCreative under the model IDs eleven_v4 and eleven_v4_turbo.

Read that list as a product statement rather than a scoreboard. AssemblyAI built a TTS endpoint for one job — reading a turn of dialogue into a phone line fast, correctly, and at a fixed price — and deliberately left out the studio controls. ElevenLabs kept building a voice platform. If your product is dubbing, narration, or a brand voice that has to be trained and governed, the six-language endpoint is not in the running regardless of price.

The latency numbers disagree by a factor of three, and both are honest

This is the part of the matchup most coverage gets wrong, in both directions. AssemblyAI's launch post quotes 90 ms to first audio in production and a 47 ms median time to first byte in its own lab test, described as fastest of six models. Its self-reported head-to-head table puts 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, and ElevenLabs v4 at 612 ms. ElevenLabs' own documentation claims roughly 150 ms median time to first speech and about 100 ms median inference latency for Turbo.

Then Cekura, an independent testing firm, measured AssemblyAI-TTS-1 at 159 ms median streaming latency, eighth of the nineteen models it timed, with 171 ms at the median when the full text is sent and 296 ms at p90. That is not a contradiction of the vendor's 47 ms. It is a different stopwatch: AssemblyAI times request to first byte, Cekura times from the last word sent to the first audio out. A voice agent negotiating a turn cares about the second measurement more, because the LLM has to finish generating the sentence first.

So the honest reading is that on the vendor's own methodology AssemblyAI-TTS-1 is the fastest thing in the comparison and ElevenLabs v4 is the slowest by a wide margin, while on an independent agent-shaped methodology AssemblyAI-TTS-1 lands mid-pack and both its claims and ElevenLabs' claims should be treated as unverified for your telephony path. The one latency comparison that survives scrutiny is against Flash v2.5, which is the ElevenLabs SKU that exists to win this argument and which third parties have measured between roughly 185 and 197 ms in production. AssemblyAI-TTS-1 at 159 ms is faster than that, at 2.7× lower price. It is the only place in this matchup where cheaper and faster are the same answer.

Where AssemblyAI-TTS-1 breaks first

Not on speed and not on price. On the long tail of things a real caller says.

Cekura's per-category results put AssemblyAI-TTS-1 first on phone numbers and codes at 91.7%, fourth on naturalness at 4.50, and twenty-second of twenty-seven on contextual pronunciation at 70.8%. Artificial Analysis likewise scores it 81.7% overall on pronunciation robustness — sixteenth of the models it scored — with the weakest sub-score on expanding shorthand at 73.8%. Coval's leaderboard puts its word error rate at 1.7%, three-tenths of a point behind the leader and ahead of Eleven v4 Turbo and Cartesia Sonic 3.6 at 1.8% and Deepgram Aura-2 at 1.9%.

Put those side by side and a specific failure mode appears: this model is very good at reading a number aloud and noticeably weaker at deciding how an abbreviation, an address, or an account code should be spoken. That is precisely the class of error a pronunciation dictionary fixes, and AssemblyAI-TTS-1 does not ship one yet. If your agent reads order IDs, you are buying a model whose best feature covers it and whose worst feature is the thing you cannot configure around.

ElevenLabs' answer to the same problem is not better scores, it is finer control: dictionaries, tags, speed, and a professional clone trained on your own reference. You are paying $65 more per million characters for the ability to fix the tail yourself.

The other half of the agent is still on a different bill

A TTS model never runs alone in a voice agent. It runs behind a language model that writes the sentence, and in most stacks that language model is the larger line item and the one with the more volatile vendor relationship. That is the layer OrcaRouter covers: one API for 200-plus models at provider list price with 0% markup, automatic failover when a provider degrades, a routing DSL for composing models into a single call, and model fusion when one answer is not good enough.

Two honest caveats. First, AssemblyAI-TTS-1 is not on OrcaRouter's catalogue today — the speech routes we carry are OpenAI's tts-1, tts-1-hd and gpt-4o-mini-tts and Google's Gemini TTS previews — so this is not a pitch to buy the model here. Second, and more useful: because markup is zero and provider list prices pass through, a vendor price cut on whichever TTS model you do route shows up on your bill the same day, which is the mechanism that makes the $80-to-$40 tier shuffle above worth watching rather than just reading about.

What to do on Monday

Three questions decide this faster than any benchmark.

Does your agent need to speak a language outside the six AssemblyAI-TTS-1 launched with? Then the comparison is over, and it is over on coverage, not quality. Wait for the language list to grow or pay ElevenLabs.

Does your agent read identifiers, codes, or addresses aloud to a customer? Then test AssemblyAI-TTS-1 on your own strings before you commit to it, because the independent pronunciation numbers are the weakest part of its record and the configuration escape hatch does not exist yet. The vendor's 93.7% entity accuracy on its own 580-item set — against 93.6% for Eleven v4 and 68.0% for Flash v2.5, vendor-reported and unreproduced — says the opposite of what Cekura's contextual score says. One of those two is measuring your workload and you do not know which.

Is the voice yours, trained and governed? Then ElevenLabs, without much argument. Instant cloning from 24 seconds produces a usable voice ID; it does not produce a brand asset you can defend.

If the answers are six languages, no, and no — a phone agent at scale reading ordinary conversational turns — then AssemblyAI-TTS-1 at $15 with unlimited concurrency and no concurrency fees is the cheaper and, on the independent latency measurement, faster option, and the quality gap is 107 Elo of blind preference you are not paying $6,750 a month to close. Put it behind a failover path, run it against your own call recordings for a month, and keep Flash v2.5 wired up as the fallback for the day the tail of your corpus turns out to be full of abbreviations.

That is the whole matchup. One vendor is selling a ceiling and the other is selling a floor, both priced accordingly, and the only expensive mistake available here is buying the ceiling for a workload that never reaches it.