📖 8 min read
Updated July 3, 2026. Voice AI is its own pricing universe. It is also 2–3x more expensive than chat per useful minute, mostly because audio tokens are priced higher than text tokens and conversations take longer than people think. The OpenAI Realtime API and its sibling transcription models are the most capable voice stack in 2026, but the rate card is structured differently from the chat rate card and most teams miss it until the bill arrives.
Here is the full Realtime rate card, what 1,000 minutes of voice actually costs across four common use cases, and when you should reach for Realtime vs batch Whisper.
Why Realtime is its own pricing universe
Three reasons voice costs more per “useful minute” than text:
- Audio tokens are denser than text tokens. One minute of 16kHz mono PCM audio ≈ 960 audio tokens. One minute of natural speech at 150 words/min ≈ 200 text tokens. The audio-to-text ratio is roughly 5x, but audio tokens are priced 5–15x higher than text tokens per million.
- Turn-taking is harder than you think. A 60-second voice message usually produces a 30–45 second exchange once you account for the model thinking, the user pausing, and the system detecting end-of-turn. You are billed for the thinking time and the pauses too.
- Real-time constraints prevent batching. The whole point of Realtime is low-latency streaming. You cannot move voice traffic to the Batch API for a 50% discount. Realtime is a premium product at a premium price.
The full Realtime rate card (July 2026)
Five models to know. Each priced per 1M tokens (or per minute for the dedicated audio models). All prices are list rate, US region.
gpt-realtime-2 — the flagship multimodal Realtime model
| Modality | Input | Cached input | Output |
|---|---|---|---|
| Audio | $32.00 | $0.40 | $64.00 |
| Text | $4.00 | $0.40 | $24.00 |
| Image | $5.00 | $0.50 | — (no image output) |
This is the model you use for full-duplex voice conversations where the model speaks, listens, and reasons. Audio output is the killer line at $64/1M — it dominates Realtime bills.
gpt-realtime-translate — dedicated speech-to-speech translation
| Pricing unit | Rate |
|---|---|
| Per minute of input audio | $0.034 |
Specialized for real-time translation across 60+ languages. Per-minute pricing, no token math needed.
gpt-realtime-whisper — dedicated streaming transcription
| Pricing unit | Rate |
|---|---|
| Per minute of input audio | $0.017 |
The cheapest dedicated transcription model in the Realtime family. Half the price of gpt-realtime-translate because it only transcribes — no translation, no generation.
gpt-4o-transcribe and gpt-4o-mini-transcribe — the async transcription workhorses
| Model | Input per 1M tokens | Output per 1M tokens | Approx cost per minute |
|---|---|---|---|
| gpt-4o-transcribe | $2.50 | $10.00 | $0.006 |
| gpt-4o-mini-transcribe | $1.25 | $5.00 | $0.003 |
These are the async transcription models. Same Whisper-grade accuracy as gpt-realtime-whisper but priced per-token. Best for batch transcription of call recordings, voicemails, and podcast archives.
One quick example: 100 minutes of podcast transcription
To make the per-minute math concrete: a 100-minute podcast recorded at 16kHz mono. One minute ≈ 960 audio tokens, so 100 minutes ≈ 96,000 audio tokens (call it 96K for clean numbers).
| Model | Tokens | Cost |
|---|---|---|
| gpt-4o-transcribe | 96K input | $0.24 |
| gpt-4o-mini-transcribe | 96K input | $0.12 |
| gpt-realtime-whisper | 100 min × $0.017 | $1.70 |
| gpt-realtime-2 audio (transcribe-only path) | 96K input audio | $3.07 |
For a podcast you publish once and never re-touch, gpt-4o-mini-transcribe at $0.12 is unbeatable. The same 100 minutes on gpt-realtime-2 would cost 25x more — and the only thing it would buy you is lower latency, which a podcast does not need.
What 1,000 minutes of voice actually costs
Four use cases. Same volume (1,000 minutes of audio per month). Different architectures. Different bills.
Use case 1: Live voice agent (gpt-realtime-2, full duplex)
A customer-facing voice agent. The model speaks, listens, and reasons in real time. Average exchange: 60% audio input, 40% audio output (the model talks more than the user). 1,000 minutes of audio = ~600 min input + ~400 min output.
| Line | Tokens | Rate | Cost |
|---|---|---|---|
| Audio input | 576M tokens | $32/1M | $18,432 |
| Audio output | 384M tokens | $64/1M | $24,576 |
| Total | 960M | — | $43,008 |
$43 per minute of live voice agent. This is the most expensive way to build with the Realtime API. It is also the only way to build a true full-duplex conversational agent in 2026.
Use case 2: Voice agent with prompt caching (gpt-realtime-2, 80% cache hit)
The same voice agent, but with the 8K-token system prompt cached. Cache hit rate: 80% on input.
| Line | Tokens | Rate | Cost |
|---|---|---|---|
| Audio input (cached reads) | 460.8M tokens | $0.40/1M | $184 |
| Audio input (fresh) | 115.2M tokens | $32/1M | $3,686 |
| Audio output | 384M tokens | $64/1M | $24,576 |
| Total | 960M | — | $28,446 |
Caching saves $14,562/month on the input side. Audio output still dominates the bill — at $64/1M there is no caching relief on what the model says back.
Use case 3: Real-time translation (gpt-realtime-translate)
Translation use case. 1,000 minutes of source audio.
| Line | Minutes | Rate | Cost |
|---|---|---|---|
| Audio input (translation) | 1,000 | $0.034/min | $34 |
$0.034 per minute of translation. 1,260x cheaper than the live voice agent. Translation is also a much narrower product — it is not generating free-form responses.
Use case 4: Batch transcription of call recordings (gpt-4o-transcribe)
Transcribing 1,000 minutes of call recordings overnight. Async, no real-time constraints.
| Model | Audio tokens | Cost |
|---|---|---|
| gpt-4o-transcribe | 960M input | $5.76 |
| gpt-4o-mini-transcribe | 960M input | $2.88 |
| gpt-realtime-whisper | 1,000 min × $0.017 | $17.00 |
For pure transcription on recorded audio, gpt-4o-mini-transcribe at $2.88 per 1,000 minutes is the cheapest capable option. gpt-realtime-whisper at $17 is 6x more expensive and exists for streaming use cases, not batch.
Voice AI is 2-3x more expensive than chat — here’s the math
Same “1,000 minutes of use.” Same model family (gpt-realtime-2 vs gpt-4o). Here is the comparison:
| Modality | 1,000 minutes | What it would cost on chat | What it costs on Realtime | Multiplier |
|---|---|---|---|---|
| Voice (full duplex, cached) | 1,000 min ≈ 960M tokens | ~$4,800 on gpt-4o at chat rates | $28,446 on gpt-realtime-2 with caching | 5.9x |
| Transcription only | 1,000 min ≈ 960M tokens | $2,400 on gpt-4o-mini at chat rates | $5.76 on gpt-4o-transcribe | 0.0024x |
| Translation only | 1,000 min of audio | $4,800 on gpt-4o at chat rates | $34 on gpt-realtime-translate | 0.007x |
The pattern: specialized audio models are dramatically cheaper than chat models doing audio. Chat models can do audio (Realtime is one), but at a premium. Translation and transcription are narrow products with optimized token counts and they crush the cost curve.
The lesson: do not use gpt-realtime-2 for transcription. Do not use gpt-realtime-2 for translation. Use the dedicated models.
When to use Realtime vs Batch Whisper
The decision is about latency requirements, not capability. Both Realtime Whisper and gpt-4o-transcribe are based on the same Whisper architecture and have comparable word error rates. The difference is the API surface.
- Use Realtime (gpt-realtime-whisper or gpt-realtime-2) when… you need streaming transcription, sub-second latency, or the user is in the loop and waiting for live captions. Live meeting note-takers, accessibility captioning, voice agents, real-time translation.
- Use Batch Whisper (gpt-4o-transcribe via Batch API) when… you are transcribing recorded audio, voicemails, podcasts, or call archives. Async, 24-hour SLA, ~50% off list rate, no minimum latency.
- Use gpt-4o-mini-transcribe when… you are doing bulk transcription of high volumes (millions of minutes/month) and can tolerate slightly higher word error rates (typically 2–4% worse than full gpt-4o-transcribe).
FAQ
How much does OpenAI Realtime API cost per minute in 2026?
For full-duplex voice agents on gpt-realtime-2, roughly $28–$43 per minute depending on cache hit rate and audio I/O ratio. For dedicated transcription, $0.003–$0.017 per minute. For translation, $0.034 per minute. The spread is 1,000x — pick the right model.
Why is gpt-realtime-2 audio output so expensive ($64/1M)?
Because the model is generating speech token-by-token in real time, and OpenAI prices the high-quality neural audio output at a premium over text. Audio output is typically 60–70% of a Realtime bill. There is no way to cache or batch it.
Can I use prompt caching on Realtime?
Yes, on the audio input and text input sides. The cached input rate ($0.40/1M for audio, $0.40/1M for text) is 80x cheaper than fresh. Cache hit rate on a stable system prompt is typically 70–85%.
Is gpt-realtime-2 cheaper than gpt-realtime?
Yes. gpt-realtime-2 is the 2026 successor to the original gpt-realtime. The list rate is approximately 35% lower across modalities, with better latency and improved voice quality. If you are still on gpt-realtime (v1), the v2 upgrade is a free 35% bill reduction.
Does Realtime support function calling?
Yes, including tool calls during a live conversation. Tool calls add input tokens (the tool definition) and output tokens (the function call payload) but are not billed at audio rates. A typical tool call adds ~$0.002–$0.008 to the per-minute cost.
What about regional pricing differences?
OpenAI’s Realtime API is hosted in US, EU, and APAC regions. EU and APAC regions carry a 10–20% regional surcharge over the US list rate. If your data residency requirements force you to a non-US region, build that into your cost model from day one — it adds up faster than people expect.
Is Realtime pricing going to drop in 2026?
Yes, but not by as much as people hope. gpt-realtime-2 was already a 35% reduction over gpt-realtime (v1) when it launched. The next expected drop is in Q4 2026 when OpenAI typically launches the next-gen model. Don’t budget on a future price drop — build on the current rate card.
What’s the minimum viable Realtime setup to test?
10 minutes of audio through gpt-realtime-2 with caching enabled will cost you about $0.50–$0.80. That is enough to validate a voice UX before committing to a real architecture. Always prototype on the Realtime API with caching before you commit to a production model choice.
Verdict
Voice AI is 2–3x more expensive than chat per useful minute on the same vendor, and 1,000x more expensive than chat when you use the wrong model for the job. The Realtime API is the right choice when you need sub-second latency or full-duplex conversation. For everything else — transcription, translation, batch processing — the dedicated audio models are 10–1,000x cheaper.
My default architecture for a voice product in 2026: gpt-realtime-2 with prompt caching for the live agent loop, gpt-realtime-translate if I need translation, and gpt-4o-transcribe on the Batch API for any post-call transcription work. Run your projected minutes through the AI API cost calculator at the right model, not the wrong one. For broader context on how Realtime compares to chat model pricing across vendors, see GPT-5 vs Claude vs Gemini: the 2026 developer bill.
The voice pricing cheat sheet
If you remember nothing else from this article, remember this table. It is the decision tree I walk through every time someone pitches me a “voice AI” project.
| Use case | Pick this model | Approx cost per minute | Why |
|---|---|---|---|
| Live full-duplex voice agent | gpt-realtime-2 with caching | $0.028–$0.043 | Only model that does it well in 2026 |
| Live streaming captions | gpt-realtime-whisper | $0.017 | Cheapest sub-second transcription |
| Real-time translation | gpt-realtime-translate | $0.034 | Specialized, dramatically cheaper than Realtime |
| Batch transcription (recorded calls) | gpt-4o-transcribe via Batch API | $0.003 | 50% off + cheapest Whisper-grade model |
| Bulk podcast archive transcription | gpt-4o-mini-transcribe via Batch | $0.0015 | Cheapest possible, fine for 95% of podcasts |
The cost-per-minute column spans 30x from top to bottom. That is not a rounding error — it is the difference between a product that loses money and one that is profitable at scale. Pick the right row for the right workload, and voice AI is one of the most differentiated product surfaces you can ship in 2026.