ElevenLabs vs OpenAI TTS

Yes. On October 1, 2026, OpenAI deprecated tts-1, tts-1-hd and both GPT-4o mini TTS snapshots. All four are removed from the API on January 6, 2027. They are the only models the speech endpoint accepts, so any request to it that names one of them will stop working that day. ElevenLabs' speech models are not affected.

OpenAI names gpt-realtime-2.1-mini. It is not a drop-in replacement: it runs on the Realtime API over WebRTC, WebSocket or SIP rather than the speech endpoint, returns PCM or G.711 audio rather than MP3, and its voice list does not include fable, nova or onyx. Teams that only need text in and audio out can move to the Realtime API or to a dedicated TTS API such as ElevenLabs or Cartesia Sonic. Test the voice and format before January.

No. The speech endpoint converts supplied text into audio. Realtime supports a spoken conversation with audio input, conversation state and tools. Compare equivalent parts of the application before drawing conclusions about price or response time.

For anything that will run after January 6, 2027, compare against gpt-realtime-2.1-mini, because the speech-endpoint models are leaving. On ElevenLabs, v4 Turbo and Flash v2.5 target low latency, Multilingual v2 targets long-form output, and v4 and Eleven v3 target expressive delivery. Keep the model name in the evaluation results.

OpenAI offers custom voices to eligible customers through its sales process. Creation requires a consent recording and a matching audio sample. This is not a universal self-service cloning entitlement. Confirm access before planning your integration.

GPT-4o mini TTS accepts instructions about delivery, such as a cheerful or calm tone. Test those instructions with your scripts and chosen voice. Do not assume the same instructions or controls are supported by every older OpenAI speech model.

ElevenAPI lists v4 Turbo, Flash/Turbo and v3 Conversational at $0.04 per 1,000 characters, and v4, Multilingual v2 and Eleven v3 at $0.08, excluding taxes. Until January 6, 2027, GPT-4o mini TTS costs $0.60 per million text input tokens plus $12 per million audio output tokens, and tts-1 costs $15 per million characters. Its replacement, gpt-realtime-2.1-mini, bills $0.60 per million text input tokens and $20 per million audio output tokens. Estimate on the same scripts; characters and audio tokens are different units.

No. Streaming support alone does not establish which provider is faster. Playback can begin before the complete clip is ready, but response time also includes request handling, network travel and playback buffering. For an agent, include turn detection and reasoning. Measure first audible output through your real player.

Yes, if your application receives text from the language model and sends it to a separate speech service. Test text chunking and cancellation when streaming. Replacing a Realtime speech-to-speech session requires a broader design than swapping a standalone TTS request.

Verify voice rights and custom-voice access, test critical names and numbers, and check playback in the target format. OpenAI's TTS guide also requires clear disclosure that the voice is AI-generated. Include those requirements in the product flow you evaluate.