AI voice models: OpenAI vs Google Gemini TTS

AI voice models, episode 2. Previous: ElevenLabs. Next: xAI Grok voice.

This series explains one or two AI voice models per post, using the same twelve questions every time. Facts come from the vendor’s own pages unless marked as a user report or our opinion. Prices and rankings were rechecked on 27 October 2026.

This episode is not a like-for-like match, and that is the story. OpenAI is leaving standalone text-to-speech. Its deprecations page says tts-1, tts-1-hd and gpt-4o-mini-tts shut down on 6 January 2027, and the named replacement is gpt-realtime-2.1-mini, a realtime model on a different API (our story). Its voice line-up is now GPT-Live 1, a full-duplex conversation model generally available since 10 September, and gpt-realtime-2.1 and 2.1 mini, speech-to-speech models on the Realtime API. Google went the other way: on 23 September it launched two dedicated TTS models, Gemini 3.8 Flash TTS and 3.8 Flash-Lite TTS (our launch story), alongside its realtime Gemini 3.8 Live.

What each does differently. GPT-Live keeps talking while a separate backend model or agent does the reasoning and tool calls, so the voice never waits on the thinking. Gemini 3.8 TTS reads a script you give it verbatim, steered by a style prompt and inline tags, which is what you want for narration, dubbing and audiobooks. On the Artificial Analysis TTS arena (27 October), Gemini 3.8 Flash TTS is #5 (Elo 1280) and Flash-Lite #8 (1256). OpenAI’s best entry there is TTS-1 HD at #40 (1106); gpt-4o-mini-tts and GPT-Live are not listed.

OpenAI against the checklist

1. Voice cloning. Two routes. Custom voices from a recording are for eligible customers only (you contact sales): you upload a consent recording plus a sample of 10 to 60 seconds with at least 5 seconds of speech, up to 20 such voices per organisation, under a TTS Supplemental Agreement. OpenAI notes recording-based voices are not compatible with Zero Data Retention. Prompt-to-voice, open to all API customers since 16 October, designs a voice from a text description for GPT-Live only. It is free to create, 100 per organisation, and the rules forbid children’s voices and imitating named people; OpenAI says designed voices are not guaranteed to be unique (custom voices guide).

2. Languages and accents. The TTS guide lists 13 built-in voices, recommends marin and cedar, says voices “are currently optimized for English”, and follows Whisper’s language list. GPT-Live added nine voices on 16 October for German, French, Hindi, Japanese and Korean; OpenAI does not publish a total language count for it.

3. Latency and streaming. OpenAI documents no latency figure. Artificial Analysis measures time to first audio of 1.24 to 1.34 seconds for GPT-Live 1 and 1.21 seconds for Realtime 2.1 High (its measurements, not OpenAI’s). Realtime runs over WebRTC, WebSocket or SIP.

4. Emotion and style control. For TTS, an instructions field (“Speak in a cheerful and positive tone”). For GPT-Live and Realtime, the session prompt, plus the designed voice itself.

5. Output formats. TTS: MP3 (default), Opus, AAC, FLAC, WAV and raw 24 kHz 16-bit PCM. GPT-Live defaults to PCM16 at 24 kHz.

6. Price (checked 27 October, pricing). GPT-Live 1: $0.05 a minute, billed per second, with backend model and tool use billed separately. gpt-realtime-2.1: audio $32 in and $64 out per million tokens (text $4 and $24). 2.1 mini: audio $10 in and $20 out. gpt-4o-mini-tts, until shutdown: $0.60 per million text tokens in and $12 per million audio tokens out. GPT-Live has no free tier; concurrent sessions are capped at 50, 300 or 500 by plan.

7. Speech-to-speech and realtime. This is OpenAI’s whole offer now. GPT-Live is full duplex: it can listen while speaking, and hands reasoning to a backend model (OpenAI’s or your own). Realtime 2.1 is a single speech-to-speech model. On Artificial Analysis’ speech-to-speech index, GPT-Live 1 Astra scores 83.2 and Realtime 2.1 High 75.2; Artificial Analysis notes the GPT-Live result is based on one trial.

8. Long-form consistency. gpt-4o-mini-tts takes at most 2,000 input tokens per request. GPT-Live has a 128,000-token context; near 90% use it summarises and starts a replacement engine with up to 8,192 tokens of history, so very long calls are designed for. Realtime 2.1: 128,000 context, 32,000 max output.

9. Compared with previous. For TTS users the “previous” model is being removed, so the comparison is a migration: gpt-realtime-2.1-mini is not a supported model for the /v1/audio/speech endpoint, so you move to the Realtime API workflow. Among the realtime models, GPT-Live leads Realtime 2.1 by 8 points on the Artificial Analysis index.

10. What users think (user reports, not facts). Voice-agent company ThunderPhone called GPT-Live “the most natural-sounding conversationalist we have put on a phone line”, but said it “does not reliably follow instructions”. A review on eesel called it “a best-in-class voice layer that is only half of a working agent”, since the backend still does the work.

11. Licence, commercial use, open weights. Under OpenAI’s Services Agreement, you own the output, OpenAI does not train on API data unless you agree, and output may not be unique. You must tell listeners the voice is AI-generated. No open weights.

12. Our verdict (opinion). Choose GPT-Live 1 if you are building a phone or app agent and want the most natural turn-taking, and budget for the backend model on top of $3 an hour. If you use OpenAI TTS today, plan the move before January: it is a new API, not a new model name.

Google Gemini against the checklist

1. Voice cloning. Voice replication is self-serve: a clean 10 to 30 second reference clip plus a consent recording from the same adult speaker, up to 200 voices per project (guide). Google’s launch post says replication through AI Studio is not available in Israel, Texas, the EEA, the UK, Switzerland or India. Voice design from a prompt is also offered.

2. Languages and accents. Model pages list 130 languages for Flash TTS and 101 for Flash-Lite. The API guide lists 30 prebuilt voices plus “hundreds” more in an Extended Voice Library; the launch post says 2,000+.

3. Latency and streaming. Not documented as a number. Streaming is supported (L16 PCM chunks). Gemini 3.8 Live measures 1.18 seconds to first audio on Artificial Analysis.

4. Emotion and style control. A natural-language style prompt for the whole line and inline tags such as <laugh>; the text is read as a verbatim transcript. Multi-speaker output supports up to two speakers, with prebuilt voices (speech guide).

5. Output formats. WAV by default (24 kHz, mono, 16-bit), L16 PCM for streaming, and mu-law and A-law for telephony, with a sample-rate setting.

6. Price (checked 27 October, pricing, which shows a launch price through 31 December 2026). Flash TTS: $0.50 per million text tokens in, $9 per million audio tokens out, which Google states as $0.00225 per 10 seconds, about $0.0135 a minute; from 1 January 2027, $1 and $18. Flash-Lite TTS: $6 out, $0.0015 per 10 seconds (about $0.009 a minute); from 2027, $12. Gemini 3.8 Live: $3 per million audio tokens in and $12 out. A free tier exists, but Google uses free-tier content to improve its products.

7. Speech-to-speech and realtime. The TTS models do not support the Live API. Gemini 3.8 Live, labelled Stable, is the realtime model; it scores 74.5 on Artificial Analysis’ speech-to-speech index and a listener Elo of 1078 there, above GPT-Live’s 1044 to 1047. The Extended Thinking variant scores 82.6.

8. Long-form consistency. 8,192 input and 16,384 output tokens per request for both TTS models. Google does not state a duration per request. Live: 131,072 in and 65,536 out.

9. Compared with previous. Flash-Lite is Google’s named replacement for 3.1 Flash TTS Preview. On Artificial Analysis, 3.8 Flash TTS scores 1280 against 1212 for 3.1 Flash, and its audio output price fell from $20 to $9 per million tokens (rising to $18 in 2027).

10. What users think (user reports, not facts). Japanese engineer Takuya tested 58 Japanese samples: an AI judge identified “cheerful”, “whispering” and “sad” deliveries correctly every time, while Flash-Lite misread a personal name. Few hands-on reports yet in English.

11. Licence, commercial use, open weights. Google’s Gemini API terms say Google will not claim ownership of generated content, may generate similar content for others, and require users to be 18 or older. All Gemini audio carries a SynthID watermark, per the launch post. No open weights.

12. Our verdict (opinion). For anything scripted, Gemini 3.8 Flash TTS is the obvious pick at under 1.5 cents a minute until the price doubles in January, and Flash-Lite is good enough for notifications and IVR. Do not use the free tier for anything private.

Summary

#	Checklist	GPT-Live 1	Realtime 2.1 / mini	gpt-4o-mini-tts	Gemini 3.8 Flash TTS	3.8 Flash-Lite TTS	Gemini 3.8 Live
1	Cloning	custom (sales-gated, consent) and prompt-to-voice	custom voices	custom voices	replication 10 to 30 s + consent; design	same	not documented
2	Languages	not counted; 9 new voices 16 Oct	not counted	Whisper list, English-optimised	130	101	not on model page
3	Latency	AA 1.24 to 1.34 s	AA 1.21 s (High)	not documented	not documented	not documented	AA 1.18 s
4	Style	prompt, designed voice	prompt	instructions	style prompt + tags	same	prompt
5	Formats	PCM16 24 kHz	not checked	MP3, Opus, AAC, FLAC, WAV, PCM	WAV, L16, mu/A-law	same	not checked
6	Price (27 Oct)	$0.05/min + backend	$64 / $20 per 1M audio out	$12 per 1M audio out	$9 per 1M ($18 in 2027)	$6 ($12 in 2027)	$12 per 1M audio out
7	STS / realtime	full duplex	speech to speech	none	none	none	Live API
8	Long-form	128k context, handoff	128k context	2,000 input tokens	8,192 in / 16,384 out	same	131k in
9	vs previous	AA index 83.2 vs 75.2	succeeds Realtime 2	shuts down 6 Jan 2027	1280 vs 1212; $20 to $9	replaces 3.1 Preview	succeeds earlier Live
10	Users	natural, weak on instructions	few reports	few reports	strong Japanese acting	name misreading	few reports
11	Licence	you own output; closed	same	same	Google claims no ownership; closed	same	same
12	Verdict (opinion)	agents	cheaper agents	migrate	narration	budget TTS	Google agents

Listen before you choose: OpenAI’s OpenAI.fm demo and the clips on Google’s launch page. We did not generate samples of our own for this episode. For open models, see our open speech models guide.