ElevenLabs Adds Nine Text to Speech and Two Speech to Text Models to OpenRouter

Two new endpoints, one API key

Developers reach the ElevenLabs models through two endpoints: POST /api/v1/audio/speech for speech generation and POST /api/v1/audio/transcriptions for transcription. Both use an OpenRouter API key. OpenRouter says every ElevenLabs model is 50% off its list price for all customers through October 19, 8am PT.

The launch covers Eleven v4 for expressive narration, Eleven v4 Turbo for faster spoken replies, and Scribe v2 for transcripts with speaker labels. OpenRouter notes that realtime Scribe v2 over WebSocket is not part of this launch.

Which model fits which job

OpenRouter recommends starting with three models. Eleven v4 is aimed at narration, characters and product videos, where audio tags such as [whispering] and [laughing] let the script direct the delivery. Eleven v4 Turbo is meant for voice agents and live replies. According to OpenRouter, ElevenLabs built it for real-time use and it costs half the v4 rate. Scribe v2 handles transcripts, meeting notes and captions, with speaker labels and word timestamps.

The other Text to Speech models fill narrower roles:

• Eleven v3: expressive narration with audio tags, up to 5,000 characters per request.
• Eleven v3 Conversational: back-and-forth dialogue at half the v3 rate.
• Eleven Multilingual v2: stable long-form audio across many languages, with speed control, up to 10,000 characters.
• Eleven Flash v2.5: fast, low-cost multilingual speech at high volume, with 32 languages and up to 40,000 characters per request.
• Eleven Flash v2: fast English-only speech, up to 30,000 characters.
• Eleven Turbo v2.5 and v2: first-generation low-latency models, now superseded by Flash.

Eleven v4 and Eleven v4 Turbo each accept up to 10,000 characters per request. On pricing, Eleven v4, v3 and Multilingual v2 bill the full per-character rate, while v4 Turbo, v3 Conversational, Flash and Turbo bill half of it. Current prices are listed on each model page.

On the transcription side, Scribe v2 covers 90+ languages and adds audio-event tags. Scribe v2 Medical is tuned for medical terms and, according to OpenRouter, makes 35% fewer errors on clinical audio than Scribe v2. Both Scribe models share the same request shape and the same per-second price.

A walkthrough from narration to voice agent

OpenRouter's post builds a small product tour in three steps. First, Eleven v4 narrates a script. The voice parameter accepts any of ElevenLabs' 21 premade voices by name, and names are not case-sensitive. Audio tags are read as delivery directions rather than spoken, but they count toward billed characters. Because a full script can exceed the per-request limit, the example splits text on paragraph boundaries. ElevenLabs-specific settings go under provider.options.elevenlabs, where previous_text and next_text keep intonation consistent across joins and a fixed seed makes reruns more repeatable.

Second, a voice agent chains three models: Scribe v2 transcribes the question, a chat model answers, and Eleven v4 Turbo speaks the reply. The example uses openai/gpt-5-mini, but OpenRouter says any chat model in its catalog can be swapped in without touching the audio calls. The transcription response is JSON with the text and a usage object reporting audio seconds and the dollar cost.

Third, Scribe v2 turns a recorded team meeting into speaker-labelled notes. The request uploads the file as multipart, asks for verbose_json with word timestamps and sets diarize under provider.options.elevenlabs. Every word then carries a speaker, and passing num_speakers can improve the split when the headcount is known. Each word entry includes the word, start, end, confidence, speaker and speaker_label. Setting tag_audio_events adds entries such as (laughter). The turns can then go to a chat model for a summary and action items.

What it means for developers

For teams already routing text models through OpenRouter, the change puts speech generation and transcription behind the same API key and billing. A full voice loop of listening, reasoning and speaking can be built from a single platform, and the chat model in the middle can be changed freely. The 50% discount running through October 19 lowers the cost of trying the models, though OpenRouter points users to the individual model pages for current per-character prices.