Eleven v4 vs OpenAI TTS-1 HD: Two Years of Drift in One Board

OpenAI TTS-1 HD has a release date of November 6, 2023. ElevenLabs Eleven v4 has one of September 28, 2026. On Artificial Analysis' Provider Voice Arena those two dates are 216 Elo points apart — TTS-1 HD sits at rank 35 with 1,104 across 3,500 appearances at $30.00 per million characters, while Eleven v4 sits at rank 1 with 1,319 across 1,674 appearances at $80.00. Both figures carry tight intervals, ±12 and ±19, so the ordering is not in doubt. The interesting question is not which model is better, because that was settled before this article existed. It is why a 2023 endpoint still has 3,500 appearances on a board where a five-day-old model has 1,674, and what that tells you about the migration you are being asked to consider.

The scoreboard, and the one row where OpenAI wins

Five dimensions decide this matchup, and both sides of each are worth reading next to each other.

• Blind preference, vendor voices — Eleven v4 rank 1, Elo 1,319 (±19, 1,674 appearances) vs TTS-1 HD rank 35, Elo 1,104 (±12, 3,500 appearances). A 216-point gap that survives the error bars.

• Board price, per million characters — Eleven v4 $80.00 vs TTS-1 HD $30.00. OpenAI is 62% cheaper at list, and cheaper still during the ElevenLabs promotion.

• Voices — Eleven v4 documents instant cloning from ten seconds of audio plus a professional clone tier; TTS-1 HD ships a fixed voice set, nine for the tts-1 family in OpenAI's own documentation (alloy, ash, coral, echo, fable, onyx, nova, sage, shimmer) and no cloning.

• Performance direction — Eleven v4 documents inline audio tags and natural-language direction prompts; TTS-1 HD takes plain text, and OpenAI's steerable-instructions parameter belongs to its newer TTS model rather than to this one.

• Release date — Eleven v4 September 28, 2026 vs TTS-1 HD November 6, 2023.

Price is the row OpenAI wins and it is a bigger win than it looks, because the ElevenLabs promotion is temporary. At $0.022 per 1,000 characters, Eleven v4 costs $22 per million until October 12 — cheaper than TTS-1 HD outright. After October 12 it costs $80 per million against OpenAI's $30, and stays there.

Why a 2023 model has twice the sample count

That number is the most useful thing on this board and it is easy to misread. Arena appearances accumulate with traffic, and traffic accumulates with integrations that were built once and never revisited. TTS-1 HD has 3,500 appearances behind it because it has been the default speech endpoint for every team already calling OpenAI for chat completions: same key, same SDK, same billing line, one more call. Nothing in that sentence is about voice quality, and the 216-point deficit did not stop it.

This is what migration debt looks like on a ranking board. It is not a signal that the older model is competitive; it is a signal that switching costs something, and that a large number of teams have decided the something is worth more than 216 Elo points. That is a legitimate conclusion for a notification readout and a poor one for a product where the voice is what the customer hears.

The case for staying on TTS-1 HD

Three things keep it defensible, and none of them is quality.

Determinism is the first. A fixed set of nine voices produces consistent output across versions in a way that a cloning pipeline does not, and consistent output is what you want in a utility path where nobody is listening for artistry. There is no clone to drift, no reference sample to re-record, and no consent artefact to maintain.

Integration cost is the second, and it is the one the arena cannot price. Adding a second speech vendor means a new account, a new rate card, a new failure mode and a new thing to monitor. For a team already inside the OpenAI stack, that is real engineering time, and 216 Elo points do not pay for it.

Low-end volume is the third. At $30 per million characters, a small product's entire monthly voice bill is a rounding error, and there is nothing to optimise. This is the regime where the cheap answer and the lazy answer are the same answer, and that is fine.

The case for moving, and the counter-argument to it

Move when the voice is customer-facing. Ten-second cloning, inline performance direction, 90-plus languages and an input cap of 10,000 characters per request are not cosmetic differences against a fixed nine-voice set from 2023, and the arena's 216-point gap is the compressed version of a much larger capability gap. If you are building an agent, a branded assistant or anything where a listener forms an opinion of your product in the first sentence, the older endpoint is the wrong default.

The counter-argument is that "move to Eleven v4" is not the only move. OpenAI has since shipped a newer TTS model with a steerable instructions parameter, and Google's Gemini 3.8 Flash TTS sits third on the same board at 1,267 and a normalised $16.49 per million — a 163-point gain over TTS-1 HD at roughly half the board price. If your constraint is staying inside your current vendor, that is the cheaper upgrade. If it is staying inside your current cloud, it is the only one.

Where OrcaRouter actually sits in this comparison

This is the rare article in this series where we host one of the two models, so let us be precise about which. openai/tts-1-hd is live in the OrcaRouter catalogue at OpenAI's list rate passed through at 0% markup: $30 per million, model ID openai/tts-1-hd, served on the OpenAI-compatible endpoint at https://api.orcarouter.ai/v1 with a POST /v1/audio/speech route and the voice, speed and response_format parameters. Our own page reports a median latency of 2,461 ms and a 95th percentile of 4,769 ms, which is the honest reason not to put it in front of a live conversation — that is a batch-shaped number, not an agent-shaped one.

ElevenLabs Eleven v4 is not in our catalogue, and there is nothing to call through us: it is reached through ElevenLabs' own API at ElevenLabs' own rate card. What a single key changes is the shape of the migration. If the reason you have not tested the new leader is that it means a second vendor integration, the cost of finding out is one API call against a key you already hold rather than a procurement cycle. Provider list prices pass through at 0% markup, so a vendor repricing — including ElevenLabs' October 12 reversion — is live on our side the day it happens, and automatic failover means a speech path under evaluation does not have to become a production dependency while you decide.

The arithmetic that should settle it

Five million characters a month, which is a mid-size product and roughly 100 hours of speech. TTS-1 HD costs $150. Eleven v4 costs $110 during the promotional window and $400 after October 12. Gemini 3.8 Flash TTS, at the board's normalisation, costs about $82.50.

Read those four numbers back and the decision splits cleanly. During the window, the newest and best-ranked model on the board is also the second cheapest — cheaper than the 2023 endpoint it beats by 216 Elo points. After October 12 it is the most expensive option on the list by a factor of nearly five. Nothing about the models changes on that date; only the invoice does.

So: if you are evaluating this month, evaluate with the promotional rate and write the list rate into the business case. If you are shipping next quarter, plan from $0.08 per 1,000 characters and treat any comparison that quotes $0.022 as a document with an expiry date printed on it. And if you are on TTS-1 HD and the honest reason is integration inertia, the useful first step is not a migration plan — it is a single A/B on your own scripts, priced at the rate you will actually be paying in November.