Voice APIs come in three shapes: text-to-speech, speech-to-text, and realtime speech-to-speech agents. They're billed per character, per minute or per audio token, so this page converts them into terms you can compare.
For transcription, Whisper costs $0.006 / min. For text-to-speech, OpenAI's TTS-1 costs $15 / 1M chars and Google's Gemini 3.1 Flash TTS $20 / 1M audio tok out. For live voice agents, GPT-Live-1 charges $0.05 / min (voice layer) for the voice layer, while GPT-Realtime-2.1 bills audio tokens at $32 in / $64 out per 1M audio tok.
| Model | Provider | Price | Notes |
|---|---|---|---|
| GPT-Realtime-2.1 | OpenAI | $32 in / $64 out per 1M audio tok | Speech-to-speech agents with tool use |
| GPT-Live-1 | OpenAI | $0.05 / min (voice layer) | New (Sep 10, 2026), full-duplex real-time voice; billed per second, backend reasoning model billed separately |
| Grok Voice 3.0 | SpaceXAI | $0.08 / min | Real-time speech-to-speech (grok-voice-3.0); text input $0.004; replaces the deprecated grok-voice-2.0 at the same price |
| Gemini 3.1 Flash TTS | $20 / 1M audio tok out | High-quality controllable TTS |
| Model | Provider | Price | Notes |
|---|---|---|---|
| TTS-1 HD | OpenAI | $30 / 1M chars | HD text-to-speech; shuts down Jan 6, 2027 — OpenAI's replacement is gpt-realtime-2.1-mini |
| TTS-1 | OpenAI | $15 / 1M chars | Standard text-to-speech; shuts down Jan 6, 2027 — OpenAI's replacement is gpt-realtime-2.1-mini |
| Whisper | OpenAI | $0.006 / min | Transcription (speech-to-text); shuts down Feb 26, 2027 — replaced by gpt-transcribe ($0.0045/min) or gpt-live-transcribe |
| GPT-Transcribe | OpenAI | $0.0045 / min | Cheapest OpenAI transcription; recommended replacement for Whisper |
Per-minute prices (Whisper, GPT-Live-1) are the easiest to budget: multiply by call or recording length. Per-character TTS (TTS-1, TTS-1 HD) depends only on script length, so count the characters in a typical script and multiply. Gemini 3.1 Flash TTS bills audio output tokens, and Google meters audio at 25 tokens per second, so one hour of speech is 90,000 output tokens plus a small text-input charge. Realtime models bill audio tokens in both directions (GPT-Realtime-2.1 charges input audio at half its output rate), so a talkative user adds to the bill, not just a talkative agent.
GPT-Live-1 is billed per second, but its price covers only the voice layer: the backend reasoning model and any tool calls are billed separately at their own rates, so add an LLM cost per turn to your estimate. GPT-Realtime-2.1 is a single speech-to-speech model with tool use, billed on audio input and output tokens; cached audio input is charged at a small fraction of the normal rate, which matters for long sessions that resend the same system prompt. On xAI, check the exact model ID: its pricing page now lists grok-voice-3.0 for speech-to-speech and marks grok-voice-2.0 deprecated; its speech-to-speech docs say the grok-voice-latest alias now points to grok-voice-3.0, at the same price as 2.0. Pin a version if you need stable behavior.
OpenAI is retiring several older audio models. TTS-1 and TTS-1 HD shut down on January 6, 2027, with gpt-realtime-2.1-mini named as the replacement. Whisper (whisper-1) shuts down on February 26, 2027, along with gpt-4o-transcribe and gpt-4o-mini-transcribe; the replacements are gpt-transcribe and gpt-live-transcribe, and OpenAI lists gpt-transcribe at a lower per-minute rate than Whisper. Google also lists newer Gemini 3.8 Flash TTS and Flash-Lite TTS models with introductory prices that rise on January 1, 2027; even at the higher January rates, both list a lower audio-output price than 3.1 Flash TTS, so they're worth testing for new work.
If audio doesn't need to be generated live (audiobooks, course narration, pre-recorded IVR prompts), Gemini 3.1 Flash TTS on Google's Batch tier costs half the standard rate, and the model also has a free tier for testing. xAI sells standalone text-to-speech at the same per-character rate as $15 / 1M chars, plus speech-to-text priced per hour, separate from its realtime voice models.
Of the models listed here, Whisper at $0.006 / min is the cheapest. Whisper shuts down on February 26, 2027, and OpenAI lists its replacement, gpt-transcribe, at a lower per-minute rate.
TTS-1 costs $15 / 1M chars and TTS-1 HD $30 / 1M chars. Both shut down on January 6, 2027; OpenAI names gpt-realtime-2.1-mini as the replacement.
With GPT-Live-1, start at $0.05 / min (voice layer) for the voice layer, then add the backend model's token cost and any tools. GPT-Realtime-2.1 bills audio tokens ($32 in / $64 out per 1M audio tok), so its per-minute cost depends on how much each side talks.
Yes. Google's pricing page lists a free tier for Gemini 3.1 Flash TTS on standard (non-batch) requests; paid usage is $20 / 1M audio tok out, with Batch at half that.
How we choose: prices come from each vendor's official pricing page and are re-checked every week. We take no affiliate fees, so rankings aren't for sale.