01
Cheaper than the model it replaces
$0.0045 a minute against $0.006 for whisper-1 and roughly $0.006 for gpt-4o-transcribe. Price went down and quality went up, which does not happen often enough to ignore when it does.
★ THE DEFAULT · $0.27 PER HOUROpenAIModel brief
OPENAI · FILE TRANSCRIPTION · AUDIO APIGPT-Transcribe. The one most workloads should be on.
At $0.0045 a minute this undercuts the legacy whisper-1 endpoint while being materially better, and it is the model most transcription pipelines should default to. It serves both the file endpoint and a realtime transcription session, and it takes keyword hints and unstructured context, which moves accuracy on domain audio more than switching vendors usually does.
Audio APIgpt-transcribeBatch transcriptionEval pipelines
Released
May 2026
Model ID
gpt-transcribe
Audio
$0.0045 / minute
Per hour
About $0.27
Modalities
Audio, text → text
Endpoints
File + realtime session
Streaming
Supported
Tuning
Keyword + context hints
[TL;DR FOR CEO + CTO]
What to know.
- 01
It is cheaper than the model it replaces, which is unusual.
whisper-1 lists at $0.006 a minute and gpt-4o-transcribe at roughly the same. GPT-Transcribe is $0.0045 and stronger. There is no version of the argument where you keep paying more for the older model on the API, unless you need Whisper's open weights on your own hardware.
- 02
About $0.27 per hour of audio.
A thousand hours a month costs around $270. That is low enough that transcription stops being the constraint on what you can afford to process, which changes what is worth building: full-archive search, transcript-backed analytics, and per-call quality review all become reasonable rather than aspirational.
- 03
It serves both endpoints, so you can start simple.
The same model handles a file upload through v1/audio/transcriptions and a live session through v1/realtime/transcription_sessions. Most teams should start on the file path, ship, and only move to a session for the streams where someone is genuinely waiting.
- 04
Keyword hints are the highest-leverage tuning you have.
It accepts unstructured context and keyword hints for domain terms, multilingual audio, and code-switching. Feeding in product names, drug names, ticker symbols, or internal jargon typically moves accuracy on domain audio further than moving to whichever vendor currently tops a leaderboard.
- 05
Transcription is step one, not the job.
The model gives you words. Diarisation, timestamps, formatting, redaction, and the handoff into summarisation, search, or extraction are all yours to build, and that is where the effort and the risk actually sit on every project we have shipped.
[SOFTWARE DEVELOPMENT IMPACT]
What it changes for the team building with it.
Two comparisons decide most pipelines: against the legacy Whisper endpoint you are probably still calling, and against the streaming models that cost nearly four times as much.
| Dimension | vs whisper-1 | vs the streaming models |
|---|---|---|
Cost per hour | About $0.27 against $0.36 for whisper-1, so the migration pays for itself immediately rather than needing a business case. gpt-4o-transcribe sits at roughly $0.006 a minute and is also undercut. | About $0.27 against roughly $1.02 for gpt-realtime-whisper or gpt-live-transcribe. That gap is the whole architectural decision: pay 3.8 times for latency, or do not. |
Accuracy on domain audio | Stronger out of the box, and the keyword and context hints are the bigger change. Being able to tell a transcription model your product vocabulary was not available on the original Whisper endpoint. | A batch pass sees the whole file before deciding, where a streaming model commits with a rolling window. On hard audio the batch result generally reads better, which is why we often run both: the stream for the interface, a batch pass for the record. |
Integration effort | A model-string change on the file endpoint for most codebases. This is the cheapest migration in this entire category, and there is no reason to defer it. | A file upload is a request and a response. A transcription session is a persistent connection with reconnection, partial handling, and revision logic. Start here and only take on that complexity where a user is watching the words appear. |
What it unlocks | At this price, transcribing everything becomes defensible. Full-archive search, quality review across every call rather than a sample, and transcript-backed analytics all move from too expensive to obvious. | Streaming unlocks live captions, in-call assistance, and voice interfaces. If none of those are on your roadmap, the streaming tier is buying you a property no user will ever perceive. |
We build the whole path: transcription, diarisation, timestamps, redaction where the data needs it, and a clean handoff into whatever reads the text next. The model choice is the cheapest decision in that chain, and the one people spend the most time on.
[WHAT IS NEW]
The features that ship with it.
02
One model, two endpoints
It serves v1/audio/transcriptions for files and v1/realtime/transcription_sessions for live audio. You can prototype on the file path and move a specific stream to a session later without changing models or re-running the vendor evaluation.
03
Keyword hints and unstructured context
Pass domain terms, product names, and freeform context to improve recognition of vocabulary the model has never seen. On specialist audio this is the single most effective lever available, and it is cheaper and more reversible than switching vendors.
04
Multiple language hints
Handles multilingual audio and code-switching with explicit language hints, which is the case that breaks most naive transcription pipelines. Anyone processing calls in a bilingual market should test this specifically.
05
Streaming on the cheap tier
It supports streaming even at $0.0045 a minute. The dedicated streaming models buy lower latency and, on gpt-live-transcribe, an explicit latency and accuracy dial, but streaming itself is not gated behind the higher price.
06
Rate limits that scale to real volume
500 to 30,000 requests per minute depending on tier. For a batch archive job the practical constraint is usually your own concurrency and storage rather than the API, which is a pleasant change from the realtime side.
[THE SPEC]
Everything an integration depends on.
The limits, the endpoints, and the lines that decide whether this model fits your transport and your budget. Kept here so nobody has to reconstruct it from three vendor pages.
| Model ID | gpt-transcribe |
|---|---|
| Input | Audio and text |
| Output | Text |
| Endpoints | v1/audio/transcriptions and v1/realtime/transcription_sessions |
| Streaming | Supported on both paths |
| Tuning surface | Unstructured context, keyword hints, and multiple language hints for domain terms, multilingual audio, and code-switching |
| Pricing | $0.0045 per minute of audio |
| Rate limits | 500 to 30,000 requests per minute depending on usage tier |
| Streaming siblings | gpt-realtime-whisper and gpt-live-transcribe, both $0.017 per minute |
| Superseded | whisper-1 at $0.006 per minute and gpt-4o-transcribe at about $0.006 per minute |
[WHICH MODEL FOR WHICH JOB]
Routing, not picking a favourite.
Where each model in the transcription lineup belongs. The split is almost always latency, not accuracy.
GPT-Transcribe
gpt-transcribe
- Meeting, call, and media transcription after the fact
- Archive processing and full-corpus search
- Feeding summarisation, extraction, and analytics
- Anything where nobody is watching the words appear
$0.0045 per minute. The default. Move off it only when a specific stream needs live output.
GPT-Live-Transcribe
gpt-live-transcribe
- Live captions with domain vocabulary
- Multilingual live audio and code-switching
- Cases where you need to trade latency against accuracy explicitly
$0.017 per minute, with a latency and accuracy dial plus keyword and language hints. Our first pick among the streaming options.
GPT-Realtime-Whisper
gpt-realtime-whisper
- Live captioning and real-time note taking
- Voice interfaces that need transcript deltas
$0.017 per minute. Whisper's streaming successor. Compare it against gpt-live-transcribe on your own audio, because they cost the same.
Whisper v1, self-hosted
whisper-1
- Audio that cannot leave your infrastructure
- Air-gapped or heavily regulated environments
- Research where you need the weights themselves
Open weights. On the hosted API it is beaten on both price and quality, so the only reason to choose it is control.
[PRICING]
What it costs to run.
Per minute of audio, at list. Transcription is one of the few AI line items you can forecast accurately before you build.
| GPT-Transcribe | $0.0045 / min | About $0.27 per hour. The default for anything not live. |
|---|---|---|
| GPT-4o Transcribe | about $0.006 / min | Prior generation. $2.50 text in and $10 out on the token path. |
| Whisper v1 | $0.006 / min | Legacy hosted endpoint. Undercut on both price and quality. |
| GPT-Live-Transcribe | $0.017 / min | About $1.02 per hour. Streaming with a latency and accuracy dial. |
| GPT-Realtime-Whisper | $0.017 / min | About $1.02 per hour. Streaming transcript deltas. |
| GPT-Realtime-Translate | $0.034 / min | Speech to speech translation, 70+ input languages. |
The arithmetic worth doing before you build: take your monthly audio hours, split them by whether a human is waiting on the words, and price each half at the right rate. A thousand hours entirely on the streaming tier is roughly $1,020 a month. The same thousand hours split realistically, with maybe a tenth genuinely live, is closer to $350. That difference is usually larger than any saving available from switching vendors.
[RISK + GOVERNANCE]
What to watch.
Cheap transcription tempts you to transcribe everything.
At $0.27 an hour the cost stops being the brake, which means the decision about what to record and retain has to be made deliberately rather than by budget. Every hour you transcribe creates a searchable text record of a conversation, and that record has a retention policy whether you have written one or not.
Consent rules vary and they bind the pipeline.
Recording and transcription consent differs by jurisdiction, with some requiring every party to agree. That shapes what you can capture, what you must disclose, and how long you may keep it. Settle it before the pipeline exists, because unwinding a transcript store afterwards is expensive and awkward.
Accuracy is uneven across speakers, and that has consequences.
Word error rate is systematically worse on accented speech, non-native speakers, and code-switching. If transcripts feed quality scoring, compliance review, or any decision about a person, the error distribution is a fairness question. Build an eval set that deliberately covers the speakers your headline accuracy number averages away, and keep a human review path.
[OUR TAKE]
How we are using it.
01
If you are still on whisper-1, move today.
It is a model-string change, it costs less, and it performs better. This is the rare migration with no trade-off to weigh, and we have not found a reason to leave a hosted whisper-1 call in place on any codebase we have picked up this year.
02
Most streaming pipelines are streaming out of habit.
The first thing we do on a transcription review is separate the streams a human actually watches from the ones nobody does. On every project so far, the second group has been larger, and moving it here has funded better work elsewhere in the pipeline.
03
Try hints before you try another vendor.
Passing your product vocabulary and a paragraph of context typically beats a vendor migration on domain audio, and it takes an afternoon instead of a quarter. We only start comparing providers once the hints have been exhausted.
04
Two passes is a legitimate design.
Where the transcript is both an interface and a record, run the streaming model for what the user sees and a batch pass afterwards for what gets stored. At $0.0045 a minute the second pass barely registers on the bill and produces the better text.
05
The model is not the project.
Every transcription engagement we have run spent most of its effort on diarisation, timestamp alignment, redaction, formatting, and the handoff into search or summarisation. Choosing the model is an afternoon. Building the path that makes the words useful is the work.
[METHODOLOGY · K-FRAMEWORK]Integrated through the
Integrated through the
K-Framework.
Every model we integrate runs through the same operating system. Three pillars, sixteen layers, one Compound Growth Loop. The methodology that keeps AI work from rotting after the first ship.
Read the K-Framework01
Foundations
Direct API integration with the model. No LangChain, no orchestration vendor, no agent framework built on quicksand. Typed contracts, the same way we wire up Postgres.
02
Amplification
An eval suite built from your real tasks gates every prompt and model change. Quality is measured before it ships, not vibed in a demo.
03
Judgment
Governance, audit, and oversight wired in from day one. Who called what, with which prompt version, at what cost. Your auditors get answers, not screenshots.
[OBSERVABILITY]
Observability your team can read.
A model in production without observability is roulette. We instrument every integration so engineering and finance can see the same numbers, and so a regression at 3am surfaces before a customer opens a ticket.
Instrumented
Cost per call
Tokens in, tokens out, dollars spent. Sliced by feature, tenant, and route. Budgets enforced where it matters.
Instrumented
Latency p50 / p95 / p99
Real distributions, not averages. We know which routes are slow, and why.
Instrumented
Eval pass rates
The same eval suite that gates a release runs continuously in production. A regression on real traffic surfaces fast.
Instrumented
Prompt + completion logs
PII scrubbed at the proxy, shipped to your SIEM. Retention controls match your compliance window.
Dashboards your team owns, not ours. At handoff you get the queries, the alerts, and the runbook. We are not in the path to read your metrics.
[COMMON QUESTIONS]
Questions we are getting asked.
- Should we still use whisper-1?
- On the hosted API, no. GPT-Transcribe costs $0.0045 a minute against $0.006 and is better, so there is no trade-off to weigh. The one genuine reason to stay with Whisper is that its weights are open and can run inside your own infrastructure, which matters when the audio is not allowed to leave.
- Does it do diarisation and timestamps?
- Not on its own. The model returns text, and speaker labels, timestamps, formatting, and redaction are pipeline work. If you would rather buy that than build it, Gemini 3.5 Transcribe carries diarisation and word timestamps in the model at $0.003 per minute of audio, which is worth pricing against GPT-Transcribe plus a diarisation stage before you commit to building one.
- How do we improve accuracy on our jargon?
- Use the keyword hints and unstructured context the model accepts. Pass product names, internal terminology, drug names, ticker symbols, and a short description of the domain. On specialist audio this reliably beats switching to whichever vendor currently leads a public leaderboard, and it takes an afternoon to try.
- Can it handle bilingual calls?
- It takes multiple language hints, which is aimed exactly at multilingual audio and code-switching. That is the case where naive transcription pipelines fail worst, so if you operate in a bilingual market, test it specifically and build your eval set around those calls rather than around your cleanest ones.
- When do we need the streaming models instead?
- When a person is watching the words appear: live captions, in-call assistance, real-time note taking, voice interfaces. The streaming tier costs 3.8 times as much, and that premium buys latency alone. If the transcript is read after the fact, by a human or a model, it buys you nothing.
- What does this cost at scale?
- About $0.27 per hour of audio, so a thousand hours a month is roughly $270 and ten thousand is roughly $2,700. That is usually small enough that the storage, the downstream LLM calls, and the engineering all cost more than the transcription. It is worth metering per pipeline anyway, so a runaway job shows up as a line item rather than a surprise.
DIRECT INTEGRATION · NO FRAMEWORKWant GPT-Transcribe
Want GPT-Transcribe
in your product?
Eval suite at handoff, full source ownership. We integrate against the model API the same way we integrate against Postgres, and route by task at runtime. Sized to your scope.