Latest in AI
HeyGen Voice Turbo API: Free Access, Pricing and Limits
HeyGen's faster speech model has a free endpoint through November 30, but production reliability and voice controls need a closer look.
HeyGen made Voice Turbo available through its API, with a separate free endpoint running through November 30, 2026. Announced on October 21, it gives developers access to the faster mode of HeyGen's speech model. The free option has shared capacity, so a zero price does not guarantee a response.
The release is worth separating from HeyGen's avatar products. You can generate speech on its own, using a public voice or an instant voice clone, without creating a video.
HeyGen Voice Turbo pricing
HeyGen's launch page lists three choices. Prices below are per million characters, not per million tokens.
| Model | Listed price | Supported voices |
|---|---|---|
| Turbo Free | $0 through November 30, 2026 | Public and instant |
| Turbo | $4 per million characters | Public and instant |
| HeyGen Voice | $15 per million characters | Public, instant and professional |
The model identifiers are heygen-voice-1-turbo-free, heygen-voice-1-turbo and heygen-voice-1 respectively. Choosing the free identifier matters: leaving the model unspecified uses the paid default.
Free access needs an account and API key, but HeyGen says it does not require a credit card. Instant voice creation is also free during the preview, within the workspace's applicable limits. Professional clones remain a different workflow and do not work with either Turbo model.
How the speech API works
The model documentation distinguishes instant cloning from professional training. An instant clone uses a single recording, taking its first three minutes, and is usually ready within seconds. Professional cloning uses a larger recording set and a training step.
Both speech delivery options use the same request fields. The completed endpoint, POST /v3/models/audio/tts, returns a WAV download URL. The streaming endpoint, POST /v3/models/audio/tts/stream, returns audio parts through Server-Sent Events. Streaming therefore needs a client that handles arriving events, rather than waiting for a finished file.
The speech API guide limits text to 5,000 characters per request. It documents a default allowance of 30 requests per minute per workspace member for each endpoint, while noting that account limits can differ.
Free capacity is a second limit. Requests can return HTTP 429 even below the account's allowance. The free endpoint does not silently convert a failed request into a paid one; using paid Turbo requires a new request with that model selected.
That is a useful billing boundary. A prototype can remain free when demand spikes, but the application still needs to handle the missing audio rather than assuming payment will rescue the request.
What the benchmark actually shows
At the time of checking, Artificial Analysis' Controlled Voice leaderboard placed HeyGen Voice Turbo first at 1,202 Elo and HeyGen Voice second at 1,195. The evaluation uses the same eight cloned voices across models, four US and four UK, with listeners choosing between samples.
The confidence intervals overlap, and the page gives both HeyGen models a possible rank range of one to three. The sensible reading is strong performance in this English controlled-voice comparison, not a conclusive win for every language, accent or production script.
There is also a pricing discrepancy worth noticing: Artificial Analysis lists the standard HeyGen Voice model at $30 per million characters, while HeyGen's launch table lists $15. The table above follows HeyGen's own current announcement. Confirm the applicable billing rate before using the standard model at volume.
Where the free endpoint fits
Turbo Free makes sense for evaluating a voice, building a demonstration or generating small amounts of speech that can tolerate retries. HeyGen does not guarantee its availability, response time or time to first audio. Those are important omissions for a conversational product, where a long silence can break the interaction even if the eventual voice sounds good.
The voice controls deserve attention too. Instant voices support expressiveness adjustment, while the documented speed, pitch and seed controls belong to professional voices. A valid request does not necessarily mean every supplied control changed the output.
For a team adding narration to an application, the first evaluation should therefore include the actual scripts and delivery requirements: names, long sentences, pronunciation and the delay before usable audio arrives. A listening leaderboard answers only part of that question.
The useful economics are straightforward. Free access lowers the cost of deciding whether the voice fits. The production decision still depends on whether the paid path gives the application the response behavior and controls it needs.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help AI teams turn technical launches and product updates into clear content that developers and founders read and share.