HeyGen Voice Turbo Takes the Controlled Voice Crown From HeyGen Voice — and Cuts the Arena Price 87%

Twelve days after HeyGen Voice reached number one on Artificial Analysis's Controlled Voice Arena, HeyGen has replaced it there with HeyGen Voice Turbo, a faster and far cheaper speech model that went live on the vendor's API on 21 October 2026 under the model ID heygen-voice-1-turbo. Read on the morning of 22 October, the board puts HeyGen Voice Turbo first at 1,202 Elo and its predecessor HeyGen Voice second at 1,195, with A​libaba's Qwen-Audio-3.1-TTS-Plus third at 1,184. The headline writes itself as a succession story — a vendor beating itself — and that framing is not wrong, but it is the least useful thing on the page. The number that should change a purchasing decision sits two columns to the right, where the same leaderboard lists the new model at $4.00 per million characters against $30.00 for the one it displaced. When a leaderboard reversal like this one later turns out to rest on a stale or restated score, the correction is logged with its dates and sources in the AI Incident Archive, alongside the rest of the model and vendor failures we track.

That is an 87% reduction in listed cost attached to a model that ranks higher than the model it replaces. Both figures are Artificial Analysis conversions of HeyGen's credit pricing rather than rates HeyGen publishes, which matters and is worth holding onto — but the direction of the gap is far too large to be a conversion artefact.

What the controlled board actually says this morning

The Controlled Voice Arena holds the speaker constant: every entrant synthesises the same eight cloned reference voices, four US and four UK, and blind listeners pick the better rendition. It is the board that measures the engine rather than the vendor's demo reel, and it is the right one to read if you ship a cloned brand voice. As of 22 October 2026, the top of it reads like this.

• Rank 1 — HeyGen Voice Turbo, 1,202 Elo, ±19 at 95% confidence (interval 1,183 to 1,221), 1,685 samples, released October 2026, $4.00 per million characters.

• Rank 2 — HeyGen Voice, 1,195 Elo, ±15 (interval 1,180 to 1,210), 1,964 samples, released October 2026, $30.00 per million characters.

• Rank 3 — Qwen-Audio-3.1-TTS-Plus, 1,184 Elo, ±14 (interval 1,170 to 1,198), 2,292 samples, $19.30 per million characters.

• Rank 4 and 5 — Eleven v4 Turbo and Eleven v4, both 1,156 Elo, $40.00 and $80.00 per million characters.

• Rank 6 — Sonic 3.6, 1,151 Elo, $49.00 per million characters, on 3,193 samples — the most-voted model in the top ten.

• Rank 7 — Inworld Realtime TTS-2, 1,145 Elo, $20.80 per million characters.

Two details in that list deserve more attention than the ordering. The first is the range column: Artificial Analysis prints ranks one through three as a single band, 1-3, meaning its own statistics do not separate the top three models. The second is sample count. HeyGen Voice Turbo has 1,685 votes; Eleven v4 Turbo, two places below, has 2,593. The new model's position at the top rests on roughly two-thirds the evidence of the models it is beating, and a ±19 interval is the widest in the top five.

Seven Elo is not the story. Four dollars is

Take the two HeyGen entries as a pair, because they share an endpoint, a vendor and now a leaderboard neighbourhood. The gap between them is seven Elo points, and their confidence intervals overlap across almost their entire width: 1,183 to 1,221 against 1,180 to 1,210. Nobody should choose between these two models on quality. Whatever difference exists in cloned-voice naturalness is smaller than the uncertainty in the measurement.

The price difference is not smaller than the uncertainty. It is a factor of 7.5.

The caveat first, because it is real: HeyGen does not publish a per-character voice rate. Its public pricing page sells credits — 600 credits on one tier, 1,000 on another, larger allocations on annual plans — and never states how many characters a credit buys. The $4.00 and $30.00 figures are Artificial Analysis's own conversion, applied consistently across every vendor on the board so that models can be compared at all. They are derived numbers. What makes the comparison usable anyway is that the same conversion produced both figures: if the conversion is wrong, it is wrong for both, and a 7.5× ratio survives a constant error.

There is a third price point that does not appear on the leaderboard at all. HeyGen is running a free tier of this model, heygen-voice-1-turbo-free, through 30 November 2026. It speaks public voices and instant clones, needs no credit balance, and is explicitly documented for tests, prototypes and small amounts of speech. It is also explicitly not guaranteed: service availability, response time and time to first audio are all disclaimed, capacity is shared across every free user, and a request can return HTTP 429 rate_limit_exceeded while you are comfortably under your own 30-requests-per-minute limit. A free request never silently switches to the paid model — you have to send a new request with heygen-voice-1-turbo and accept the charge. For a production path, budget for the paid model and treat the free tier as a way to hear the engine on your own script this week.

What Turbo gives up to hit that number

The discount is not a pricing promotion dressed as a model launch. It is a narrower product, and the narrowing is documented plainly in HeyGen's speech reference.

• Voice types — heygen-voice-1-turbo accepts public voices and instant clones only. Professional clones are not supported. The default heygen-voice-1 remains the model that speaks professional voices.

• Per-request controls — Turbo accepts expressiveness_boost from 0.0 to 1.0 and nothing else in the prosody set. seed, speed, pitch_shift and pitch_variance are accepted and ignored on instant voices, including Turbo. Professional voices reject expressiveness_boost outright.

• Pauses — professional voices take <break> pauses of up to five seconds. Instant voices, Turbo included, do not. Other SSML is unsupported on either.

• Text length — 1 to 5,000 characters of plain prose per request, on both models.

• Rate limit — 30 requests per minute per workspace member by default, on both endpoints, adjustable per account.

• Output — one mono PCM16 WAV at 44.1 kHz for the synchronous endpoint; ordered base64-encoded WAV parts over server-sent events on /tts/stream. Each streamed part is a complete WAV container, so it has to be decoded individually rather than concatenated.

Read that list as a description of the customer Turbo is aimed at: someone generating short lines in a voice they cloned in seconds, at volume, where a missing pitch control is irrelevant and a 7.5× rate difference is not. If your pipeline depends on a twenty-minute professional clone and reproducible seeds, Turbo is not a cheaper version of your current model — it is a different model that cannot do your job.

The claim HeyGen is making in its own documentation

HeyGen's Text to Speech reference states that its API defaults — similarity: "extra_high" for cloning and expressiveness_boost: 1.0 for speech — are the configuration that "took first place on the Artificial Analysis Controlled Voice leaderboard." That is a vendor claim about an independent board, and it happens to be checkable: the board does show a HeyGen model in first place. The claim is also narrower than it sounds. It describes one configuration of one model on one of Artificial Analysis's two speech boards, and it says nothing about the leaderboard position of the vendor's voice catalogue as a whole.

Note the mismatch worth knowing about before you copy the defaults: HeyGen Studio ships similarity: "max" and expressiveness_boost: 0.5, while the API defaults to the higher-similarity, maximum-expressiveness pair. The leaderboard result was produced on the API defaults. If you tune for Studio-like warmth, you are no longer running the configuration the Elo belongs to.

Three things this leaderboard does not measure

First, HeyGen has no entry on the Provider Voice Arena at all — the companion board where each vendor uses its own native voices and 98 models are ranked. That board is led by Eleven v4 Turbo at 1,335, with Inworld Realtime TTS-2 sixth at 1,267. Artificial Analysis omits models that lack enough votes, so absence is not a verdict on quality; it does mean that no independent number exists for how HeyGen's own catalogue sounds to a general audience. If you are choosing among stock narrator voices rather than cloning, this article's central number does not apply to you.

Second, the controlled board runs per language, and HeyGen Voice Turbo appears at the top of the English boards and on the German board — but is absent from the Japanese, Chinese, French, Spanish, Portuguese, Vietnamese, Arabic and Hindi tabs. Given that HeyGen documents 42 base language codes for cloning and speech, the likeliest explanation is vote volume rather than capability. Still, an English-language crown is an English-language crown, and anyone shipping multilingual narration should read the relevant tab before assuming the result travels.

Third, the arena scores audio, not the cloning step that produced the voice. Whether an instant clone from a few seconds of reference audio preserves a specific speaker is a separate question, measured by nobody here.

What happens on 1 December

The most concrete open question is not whether Turbo holds first place — a seven-point lead inside overlapping intervals could reverse on the next scoring pass without anything changing at HeyGen. It is what the model costs after the free tier closes on 30 November 2026.

Two paths are plausible. The $4.00 conversion could be the steady-state price of a deliberately cheap instant-voice model, in which case the controlled board now has a top-ranked entry at a fifth of the field's median. Or Turbo Free could be an acquisition channel and the paid rate could move once it ends. HeyGen says it will notify users before the offer changes, which is the only commitment on the record. Until a vendor rate card exists, every price here is an inference — a good one, consistently applied, and still an inference.

For teams already running a voice pipeline, the practical move this week is small and cheap: render twenty of your own lines through heygen-voice-1-turbo-free with an instant clone, and render the same twenty through whatever you use today. The arena has told you the engines are close on eight reference voices. It cannot tell you whether Turbo handles your product names, your numbers and your pauses. If you also call language models or one of the hosted speech models in OrcaRouter's catalogue, those can sit behind a single key with automatic failover, which keeps the experiment off your production path — though HeyGen Voice Turbo itself is not routed there, so the comparison run has to go through HeyGen's own API.

The succession story, in the end, is the least interesting reading. A vendor improving its own model by seven Elo in twelve days is ordinary progress. A vendor shipping its best cloned-voice engine at roughly one-seventh the listed cost of the previous one is a change in what a voice layer is allowed to cost — and that is the sentence to carry into a budget conversation.