ChatGPT Realtime Audio

by OpenAI · launched 2026-05-08
NewOpenAI shipped three GPT-Realtime audio models on 2026-05-08:
All three are developer APIs requiring an OpenAI account and API key.
This page is the honest comparison vs Whipscribe — last price-check 2026-05-08.
TL;DR

If you're a developer building a voice agent and you want GPT-5 reasoning on the audio path with a single endpoint — Realtime-2 is the right tool, and Whipscribe doesn't compete on that.

If you want a transcript — paste a URL, drop a file, or hit an MCP/REST endpoint — your first Whipscribe transcript in the web app is free, at any length (API keys draw on account credit). After that, credit packs start at $8 for 1,000 minutes ($0.008/min, already under Realtime-Whisper's $0.017/min), and bigger packs bring the rate down: $12 buys 2,000 minutes (effective $0.36/hr, about a third of ChatGPT Realtime's ~$1.02/hr). Every pack is one-time payment, credits never expire, no subscription. Preview any transcript instantly with no signup. 99 languages, speaker labels included, no audio sent to OpenAI servers.

If you care about privacy — Whipscribe runs self-hosted faster-whisper / whisperX with no third-party AI calls on the audio path; ChatGPT Realtime sends every audio stream to OpenAI's US servers, retained per OpenAI's data-usage policy.

The three new models

GPT-Realtime-2

$32 / 1M audio input · $64 / 1M audio output ($0.40 cached input)

Voice agent with GPT-5-class reasoning. Carries multi-turn dialogue, calls tools, executes actions while the user is still talking. The flagship — and the most expensive of the three.

GPT-Realtime-Translate

$0.034 / minute (~$2.04/hr)

Live speech-to-speech translation. 70 input languages → 13 output languages, low-latency streaming. Aimed at meeting bots, dubbing, and customer-support overlay use cases.

GPT-Realtime-Whisper

$0.017 / minute (~$1.02/hr)

Streaming speech-to-text. Same Whisper family OpenAI has shipped before, now exposed as a real-time stream so partial transcripts appear as the speaker talks. The closest direct competitor to Whipscribe's transcription API.

Pricing pulled from OpenAI's launch announcement on 2026-05-08. Audio billing is rounded to the nearest second; Realtime-2 token math depends on input length.

At a glance

Category
Voice AI · streaming STT
Launched
2026-05-08
Cheapest tier
$0.017/min (Whisper)
Languages
99 (Whisper) · 70→13 (Translate)
Streaming
Yes — real-time
Free tier
None — pay from token zero
Account required
Yes — OpenAI + API key
Diarization
Not built-in

ChatGPT Realtime Audio vs Whipscribe

Feature ChatGPT Realtime Audio
OpenAI · 2026-05-08
Whipscribe
Neugence · privacy-first
Product category Voice AI · streaming STT API
developer-only
Transcription utility
web app · API · MCP
ChatGPT GPT · Mac desktop
Cheapest transcription rate Realtime-Whisper · $0.017/min
(~$1.02/hr)
Realtime-Translate · $0.034/min
(~$2.04/hr)
First transcript free
$0.008 / minute on the $8 / 1,000-min pack
lower on bigger packs:
$12 = 2,000 min ($0.006/min)
$24 = 5,000 min ($0.0048/min)
credits never expire · no subscription
Voice-agent / GPT-5 reasoning Yes — Realtime-2
$32 / $64 per 1M tokens
Not offered
we ship transcripts;
bring your own model
(Claude, GPT, local)
Try without signing up No
OpenAI account + API key required
Yes
instant preview, no account, no card
$0.99 opens your first transcript in full
Free tier None
billed from minute zero
First transcript free
recordings up to 3 hours
instant preview with no signup
Privacy / data residency Audio sent to OpenAI servers (US)
retention per OpenAI data policy
not HIPAA-eligible by default
Self-hosted Whisper
(our own GPU cluster)
audio never sent to OpenAI
no training on uploads
see /security + /privacy
Speaker diarization Not built-in Included
whisperX + pyannote
no extra fee
Word-level timestamps Streaming token deltas
no aligned word timings
SRT · VTT · JSON · DOCX · TXT
word-level alignment
Languages Whisper-realtime · 99
Translate · 70 → 13
99
(full Whisper coverage,
all tiers)
URL input
(YouTube, Vimeo, podcast feeds)
No
raw audio stream only
Yes
paste a link,
we pull the audio
File uploads
(.mp3, .mp4, .m4a, .wav, .mov…)
No
raw audio stream only
you write the chunking,
the resampling, the VAD,
the retry logic
Yes — drag many files
up to 10 hours / 5 GB per file
(up to 3 hours before you buy credits)
parallel jobs
auto-resume on tab close
any container we can decode
Searchable transcript library None
token deltas in a stream;
no archive, no search,
no folders, no share links —
you build all of it
Full library
every transcript saved · full-text search
folders + share links + trash
rename · delete · re-export anytime
cross-device sync via your account
Chat with the transcript
(Claude Sonnet)
No
token feed only
"summarize my meeting" =
your code, your prompt,
your model bill
Built-in
Claude Sonnet on every transcript
"summarize the decisions",
"pull every action item",
"what did Sarah say about churn?"
private · saved to your account
Live transcription Yes
real-time streaming model
Live Meeting Notes (beta)
streaming Whisper on web
Native integrations OpenAI SDK
Realtime API endpoints
REST API
MCP server (Claude/Cursor)
ChatGPT Custom GPT
Obsidian · Mac desktop
Chrome extension
Human support Community forum + docs
API status page
Enterprise tier for ticketed support
Email a human
contact@neugence.ai
same person who built the product
typically same-day reply
no tier-gating
Pricing model Usage-metered API
no caps · no free tier
First transcript free
then one-time credit packs
$8 / 1,000 min · $12 / 2,000 min · $24 / 5,000 min
$0.99 unlocks your first transcript
Credits never expire · no auto-renew
Privacy in plain English. Whipscribe is privacy-first by design: your audio hits our self-hosted Whisper / whisperX cluster and never leaves it for an OpenAI, AssemblyAI, or Deepgram round-trip. We don’t train on uploads. Anonymous transcripts are auto-deleted on a short clock; signed-in transcripts stay in your library, deletable any time. Read the full posture at /security and /privacy. With ChatGPT Realtime Audio, every audio frame is sent to OpenAI’s US servers and held under OpenAI’s data-usage policy — that may be fine for some teams, and a hard blocker for legal, medical, journalism-source, and EU-residency workflows.

What Whipscribe ships that ChatGPT Realtime doesn’t

Realtime is a developer audio API — its job ends at "stream out the bytes." Whipscribe ships the finished workflow: the transcript file, the place to keep it, the AI to talk to it, and a human you can email when something is off. Four pieces, all included on every paid pack, none of them on the Realtime side without you building them yourself.

📁 Files of any size, any format

Drop a 3-hour podcast .mp3, a Zoom .mp4, an iPhone Voice-Memo .m4a, a .wav, a .mov, an .opus — Whipscribe decodes it. Up to 10 hours / 5 GB per file (3 hours before you buy credits), and you can drag a whole folder in for parallel jobs. No chunking code, no resampler, no "audio format not supported." Realtime accepts a raw audio stream — file ingestion is on you.

🔍 A searchable library that’s yours

Every transcript you make is saved to your account, full-text searchable across the whole library — find "the meeting where we discussed pricing" in one query, six months later. Folders, share links, trash with restore, rename, re-export anytime. Realtime returns token deltas; archive + search + share is something you build.

💬 Claude Sonnet chat on every transcript

Open any transcript and ask Claude — "summarize the key decisions", "pull every action item with timestamps", "what did Sarah say about churn?" — included on every paid pack. Conversations are saved to your account and deleted with it on our server, and Anthropic doesn’t train on them. With Realtime you’d wire your own model, your own prompt, your own bill.

📨 A human who answers your email

Email contact@neugence.ai and you reach the same person who built the product — typically same-day reply, no support-tier gate. OpenAI’s Realtime API ships with docs, a forum, and an enterprise contract for ticketed support. On Whipscribe a question about your file, your bill, or a feature request goes straight to a human who can fix it.

Pricing — head to head

Workload ChatGPT Realtime Audio Whipscribe
1 hour of transcription / month ~$1.02 / hour
(60 min × $0.017,
Realtime-Whisper)
$0 for your first transcript
(free at any length)
then the $8 / 1,000-min pack
10 hours / month ~$10.20 $8 / 1,000-min pack
(covers 600 minutes · $0.008/min)
credits never expire
40 hours / month
(active podcaster / journalist)
~$40.80
(40 × $1.02/hr)
$24 / 5,000-min pack
(covers 2,400 minutes · $0.0048/min)
~3.5× cheaper per minute
2,000 minutes / month
(podcast network · research lab)
~$34.00
(2,000 min × $0.017)
$12 / 2,000-min pack
(effective $0.36/hr ·
~2.8× cheaper)
1 hour live translation ~$2.04 / hour
(60 min × $0.034,
Realtime-Translate)
First transcript free
then from $0.008/min
+ paste into Claude / DeepL
Voice-agent app
(50K interactions / month)
Token-metered
Realtime-2 · $32 / $64 per 1M tokens
Out of scope
we’re a transcription utility,
not a voice agent
Try without paying No free tier Instant preview, no signup
first transcript free with a free account

All numbers from public price pages on 2026-05-08. Realtime-2 voice-agent math depends on conversation length; the line above is illustrative only.

When ChatGPT Realtime Audio is the right call

When Whipscribe is the better fit

FAQ

Which one should I pick for my use case?

If you need a transcript file (TXT, SRT, DOCX) from a recording, an interview, a meeting, a podcast, or a YouTube link — Whipscribe is the right tool. Drop the file or paste the URL and read the transcript.

If you’re a developer building a live voice agent that needs GPT-5 reasoning, tool calls, and barge-in — ChatGPT Realtime-2 is the right tool. It’s an API, not a transcription product.

If you need live speech-to-speech translation across 70 → 13 languages with low latency — ChatGPT Realtime-Translate is built for that. Whipscribe transcribes; translation is a separate step.

Do I need to write code to use ChatGPT Realtime Audio?

Yes. All three models are developer APIs — you’ll need an OpenAI account, an API key, and code that opens an audio stream to api.openai.com. There is no web app and no upload form.

Whipscribe has a web app at whipscribe.com — paste a URL or drop a file and you get a transcript in seconds, no code. Preview any transcript instantly with no signup; $0.99 opens your first one in full, no account needed, at any length.

Can I transcribe a YouTube video or podcast URL with ChatGPT Realtime?

Not directly. ChatGPT Realtime accepts a raw audio stream — you’d need to download or capture the audio yourself and pipe it in. Whipscribe accepts a URL: paste a YouTube, Vimeo, podcast, or direct media link and the audio is fetched for you.

How much will I actually pay for typical workloads?

Realtime-Whisper is billed at $0.017 per minute, rounded to the nearest second.

  • 1 hour / month — ChatGPT Realtime ≈ $1.02 · Whipscribe $0.99 for your first transcript, then the $8 / 1,000-minute pack covers it many times over
  • 10 hours / month — ChatGPT Realtime ≈ $10.20 · Whipscribe $8 / 1,000-minute pack (covers it with room to spare)
  • 40 hours / month — ChatGPT Realtime ≈ $40.80 · Whipscribe $24 / 5,000-min pack (≈ 3.5× cheaper per minute)
  • 2,000 minutes / month — ChatGPT Realtime ≈ $34 · Whipscribe $12 / 2,000-min pack (≈ 2.8× cheaper)

Realtime-Translate at $0.034/min (≈ $2.04/hr) and Realtime-2 voice agent ($32 / $64 per 1M tokens) bill the same way — usage-metered, no monthly cap, no free tier.

Is my audio private with each tool?

ChatGPT Realtime Audio: every audio stream is sent to OpenAI’s servers in the United States and retained per OpenAI’s data-usage policy. Default API access is not HIPAA-eligible.

Whipscribe: audio is processed on Whipscribe’s own GPU cluster using self-hosted Whisper / whisperX. Audio is never sent to OpenAI or any third-party AI provider. Recordings are not used for training. The full posture is on /security and /privacy.

Does ChatGPT Realtime work for Zoom / Google Meet / Teams transcripts?

There’s no built-in meeting bot — you’d capture the meeting audio yourself, then stream it in. Whipscribe accepts uploaded recordings (mp4, m4a, mp3, wav and many more) and offers Live Meeting Notes in beta for browser-tab capture.

Which is more accurate?

Realtime-Whisper is the same Whisper family Whipscribe runs (whisper / whisperX). On clean speech the two are very close. Differences in the final transcript come from features layered on top: speaker diarization, word-level alignment, punctuation restoration, and language detection — all included on Whipscribe, not on Realtime-Whisper.

HIPAA / SOC 2 / EU data residency — what are my options?

OpenAI offers HIPAA via their Enterprise tier; default API access is not HIPAA-eligible. EU residency requires an OpenAI Enterprise contract.

Whipscribe runs on Neugence-owned infrastructure with self-hosted models and no third-party AI on the audio path. See /security for the current posture, certifications status, and how to request a DPA.

Can I use both together?

Yes — they solve different problems. A common pattern: use Whipscribe for the transcript (with diarization, timestamps, exports), then feed the text to GPT-5 / Realtime-2 to build a voice agent on top. The Whipscribe MCP server and the ChatGPT Custom GPT make that handoff one click.

Is there a free way to try either one?

ChatGPT Realtime Audio has no free tier — you pay from the first second.

Your first Whipscribe transcript is free, at any length. Preview any transcript instantly with no signup; $0.99 opens your first one in full, no account needed. After that, credit packs start at $8 for 1,000 minutes, and credits don’t expire.

Whipscribe is a managed faster-whisper + whisperX service — privacy-first, first transcript free, then credit packs from $8 / 1,000 min up to $24 / 5,000 min (credits never expire), no API key to try, 99 languages, speaker labels included.

Transcribe a file →

Cross-references: OpenAI Whisper API (older, $0.006/min) · Deepgram · AssemblyAI · all 27 tools · our security posture · our pricing.