GPT-Transcribe

GPT-Transcribe (gpt-transcribe) and GPT-Live-Transcribe (gpt-live-transcribe) are two closed-weight speech-to-text models that OpenAI added to its API on July 28, 2026.^[1][2] GPT-Transcribe handles completed audio files, streamed file transcripts, and the final transcript of a committed audio turn inside a Realtime session; GPT-Live-Transcribe is the streaming sibling, built to emit transcript deltas from live audio with tunable latency.^[3][4] This article covers both models together, because OpenAI shipped them as a pair, documents them in the same guides, and positions them jointly as the recommended starting point for any new transcription integration, with the older gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-realtime-whisper kept only for existing code.^[5] Both models accept free-form context, literal keyword hints, and a list of expected input languages, which is the feature that most clearly separates them from the Whisper line they replace.^[1][5]

Scroll sideways for more →

Background: from open-weight Whisper to API-only transcription

The direct ancestor of both models is Whisper, the encoder-decoder Transformer that OpenAI released in September 2022 and described in "Robust Speech Recognition via Large-Scale Weak Supervision" by Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever.^[8] Whisper was trained on 680,000 hours of weakly supervised multilingual and multitask audio scraped from the web, and the paper reports that the resulting models "approach" human accuracy and robustness.^[8] Just as important as the accuracy was the license: the code and weights went out under the MIT license from the openai/whisper repository, which by mid-2026 had over 106,000 stars on GitHub, while the openai/whisper-large-v3 checkpoint records roughly 5.8 million monthly downloads on Hugging Face.^[9] Whisper large-v2 (December 2022) and large-v3 (November 2023) followed, and the speech recognition field spent the next three years building on a model anyone could download.

That stopped in March 2025. On March 20, 2025 OpenAI added gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-mini-tts to the Audio API.^[6] TechCrunch reported at the time that OpenAI did not plan to release the new transcription models openly, a break with the Whisper precedent, and quoted OpenAI's claim that the models were "much improved versus Whisper" on the fabrication problem.^[10] The gpt-4o-transcribe model card still describes it as "a speech-to-text model that uses GPT-4o to transcribe audio" with a 16,000-token context window, a 2,000-token output limit, and a June 1, 2024 knowledge cutoff, all of which are attributes of a large language model rather than a classical recognizer.^[11]

The rest of the lineage filled in over the following eighteen months. gpt-4o-transcribe-diarize, a transcription model that labels who is speaking and returns a diarized_json response format, was announced on the OpenAI developer forum on October 18, 2025.^[12] In December 2025 OpenAI shipped four dated audio snapshots including gpt-4o-mini-transcribe-2025-12-15, and reported that on an internal "hallucination-with-noise" evaluation the model "produced ~90% fewer hallucinations compared to Whisper v2 and ~70% fewer compared to previous GPT-4o-transcribe models."^[13] In May 2026 the Realtime family gained gpt-realtime-2, gpt-realtime-translate, and gpt-realtime-whisper, the last of these being a streaming speech-to-text model for the GPT-Realtime transcription session type.^[6] GPT-Live-Transcribe is the model that supersedes gpt-realtime-whisper, and it is the model OpenAI benchmarked against at launch.

Endpoints, modes, and documented limits

The two models split the work by where the audio comes from rather than by size or price tier. GPT-Transcribe is the file model: send audio to /v1/audio/transcriptions and receive a transcript, optionally streaming partial text with stream=true while the file is processed.^[14] Files can be up to 25 MB, and the supported input formats are mp3, mp4, mpeg, mpga, m4a, wav, and webm; OpenAI's advice for longer recordings is to use a compressed format or split into chunks of 25 MB or less, avoiding splits in the middle of a sentence.^[14] When streaming, the model emits transcript.text.delta events and closes with a transcript.text.done event that also carries the detected languages.^[14]

GPT-Live-Transcribe runs only inside a Realtime transcription session. The client opens a session with type: "transcription", selects the model, and streams base64-encoded PCM audio (the documented example uses 24 kHz) with input_audio_buffer.append over either WebSocket for server-side pipelines or WebRTC for browser audio.^[15] Turn boundaries can be detected server-side by voice activity detection or committed manually with input_audio_buffer.commit when turn_detection is set to null.^[15] Transcript text arrives as conversation.item.input_audio_transcription.delta events, with a conversation.item.input_audio_transcription.completed event per committed item; OpenAI warns that ordering between completion events from different turns is not guaranteed and that applications should reconcile on item_id.^[15]

GPT-Transcribe can also be used inside a Realtime session, over WebSocket only, when an application specifically wants transcription to begin after a committed audio turn or needs detected-language output.^[15] In that mode it automatically uses earlier transcribed turns as context.^[5]

A latency control unique to the streaming model is audio.input.transcription.delay, which takes the values minimal, low, medium, high, and xhigh. OpenAI's documentation states plainly that "lower delay settings can produce earlier partial text" while "higher delay settings give the model more audio context before emitting text and can improve word error rate," and that the actual millisecond delay varies by configuration, so developers should benchmark rather than assume fixed timings.^[15]

The documented gaps matter as much as the features. GPT-Live-Transcribe returns no word-level timestamps, no speaker labels, and no transcription confidence scores.^[15] Word timestamps and srt/vtt subtitle output still require whisper-1, as does translation of a recording into English through /v1/audio/translations.^[5][14] Speaker-labeled transcripts still require gpt-4o-transcribe-diarize, and speaker labeling is not available in Realtime transcription sessions at all.^[14] Neither new model supports the Batch API endpoint (v1/batch), so the "batch workloads" in OpenAI's launch post means bulk offline processing of files, not the discounted Batch endpoint.^[3][1] Neither model card publishes a context window, a maximum output token count, or a knowledge cutoff.^[3][4]

Context conditioning

The three context parameters are the substance of OpenAI's "better understand context" claim. prompt takes free-form text about the recording, such as its topic or setting. keywords takes literal terms that may appear in the audio, such as product names, medications, or acronyms. languages takes a list of expected input languages, and replaces the singular language field used by older transcription models; sending both is not allowed.^[5][14] Accepted language-code formats are ISO 639-1 codes such as en, es, and fr, selected ISO 639-3 codes such as eng, spa, yue, and cmn, and regional Chinese locale codes such as zh-cn, zh-tw, and zh-hk; the API rejects unsupported or badly formatted codes.^[14][15]

OpenAI does not publish a supported-language count for either model, which is a change from the Whisper documentation that still states whisper-1 supports 98 languages.^[14] It also does not publish per-language error rates for the new models.

Two constraints are strict enough to break requests. Each keyword must sit on one line and must not contain <, >, a carriage return, or a line feed, and the API rejects the entire request when it finds one of those characters or when the prompt exceeds the model's length limit.^[14] The documentation is also unusually blunt about what keyword hints are for: "Keywords are hints, not required output. The transcript should include a keyword only when the audio contains it."^[5] That sentence exists because biasing a generative decoder toward a word list is a way to make the model write words that were never spoken.

GPT-Transcribe returns a languages array alongside the transcript, and returns an empty array when it cannot make a reliable prediction. GPT-Live-Transcribe returns no detected-language predictions.^[15]

Published accuracy figures

OpenAI released three benchmark charts with the launch, and all three compare GPT-Live-Transcribe against gpt-realtime-whisper. The company published no word error rate chart or table for GPT-Transcribe. The charts are labeled "Word Error Rate (WER)", while the accompanying forum text calls the same quantity a "transcription error rate."^[2]

Scroll sideways for more →

Source: OpenAI launch charts, July 28 to 29, 2026.^[1][2]

Two things are worth reading off that table. The gain on the multilingual Common Voice set is 0.63 percentage points, which is small; the gain on OpenAI's own real-world recording set is 2.05 points, which is larger but comes from a benchmark OpenAI does not describe publicly. And the context feature does not help the old model: gpt-realtime-whisper scores 32.2% with context against 32.4% without, essentially unchanged, while GPT-Live-Transcribe moves from 38.5% to 44.6%. The ability to use context is a property of the new model, not of the API surface.^[1]

Neither the "Real World Audio Recording" benchmark nor the "Context Aware Automatic Speech Recognition" benchmark has published composition, scoring rules, or a public test set, so both should be read as vendor-internal evaluations. Common Voice is a public multilingual corpus, but OpenAI does not list which 22 languages it used.

Aggregator coverage has circulated GPT-Transcribe figures of 19.27% on Common Voice across 22 languages and 8.98% on the real-world set, against 40.37% and 15.21% for whisper-1; those numbers appear in AlphaSignal's July 28, 2026 write-up and could not be traced to any OpenAI page, chart, or post.^[16] They are reproduced here only as that outlet's claim.

Independent evaluation

Artificial Analysis publishes an AA-WER index computed as an audio-duration-weighted average of word error rate over roughly eight hours drawn from three datasets: AA-AgentTalk at 50%, VoxPopuli-Cleaned-AA at 25%, and Earnings22-Cleaned-AA at 25%.^[17] The figures below were read from that leaderboard on July 31, 2026; it is a live ranking and the values move.

Scroll sideways for more →

Source: Artificial Analysis speech-to-text leaderboard, accessed July 31, 2026.^[17]

The German outlet The Decoder read the same leaderboard on July 29, 2026 and reported GPT Transcribe at 3.31% AA-WER, a 0.7 percentage point improvement on its year-old predecessor, and noted that the model still sits behind ElevenLabs, Google, and Mistral entries on error rate.^[18] The $4.50 per 1,000 minutes that Artificial Analysis lists is exactly OpenAI's $0.0045 per minute, which is a useful cross-check on the pricing.

Four widely used systems ship downloadable weights, which matters for anyone who cannot send audio to a third party:

Scroll sideways for more →

Parakeet TDT 0.6B v3 is a FastConformer-TDT model covering 25 European languages, and its card reports 6.34% average WER on the Open ASR Leaderboard and 1.93% on LibriSpeech test-clean.^[20] Canary Qwen 2.5B is English-only and reports a 5.63% mean WER on the same leaderboard.^[21] Both are distributed through NVIDIA NeMo.

Why LLM-backed transcription behaves differently

A classical recognizer scores acoustic evidence against a pronunciation lexicon and an n-gram or neural language model, and its output is constrained by that decoding graph. A model like GPT-Transcribe is an autoregressive sequence model that generates text conditioned on audio, so the same machinery that lets it recover a mumbled drug name from surrounding context also lets it write a plausible sentence when the acoustic evidence is thin.

Contextual biasing is the useful side of this. Passing keywords: ["premium plan", "AC-42", "billing"] and a prompt describing a support call raises the prior on those exact strings, which is why OpenAI's launch text singles out "short phrases, numbers, specialized terminology, and speech with loud background noise" as the cases that improve.^[1] Short utterances and alphanumeric strings are precisely where a classical recognizer has the least acoustic evidence to work with and where a language prior helps most. The launch video leaned on the noise case: the Japanese outlet GIGAZINE, reporting on July 29, 2026, described it as a demonstration of real-time transcription continuing while a musical instrument was played in the background.^[26]

The failure mode is hallucination. The evidence base here is Whisper's, and it is substantial. In "Careless Whisper: Speech-to-Text Hallucination Harms," presented at the 2024 ACM Conference on Fairness, Accountability, and Transparency, Allison Koenecke, Anna Seo Gyeong Choi, Katelyn X. Mei, Hilke Schellmann, and Mona Sloane found that "roughly 1% of audio transcriptions contained entire hallucinated phrases or sentences which did not exist in any form in the underlying audio," and that "38% of hallucinations include explicit harms such as perpetuating violence, making up inaccurate associations, or implying false authority."^[22] The rate was worse for speakers with aphasia, whose speech contains longer non-vocal stretches.^[22]

An Associated Press investigation by Garance Burke and Hilke Schellmann, published October 26, 2024, found the problem in deployed systems: a University of Michigan researcher found hallucinations in eight of every ten public-meeting transcriptions he inspected, a machine learning engineer found them in about half of more than 100 hours of transcripts, and a third developer found them in nearly all of 26,000 transcripts he generated.^[23] The AP also reported that more than 30,000 clinicians and 40 health systems were using a Whisper-based clinical tool from Nabla, despite OpenAI's own guidance against using Whisper in "decision-making contexts, where flaws in accuracy can lead to pronounced flaws in outcomes."^[23]

OpenAI has not published a hallucination evaluation for GPT-Transcribe or GPT-Live-Transcribe. The closest published number is for a different model: the December 2025 gpt-4o-mini-transcribe snapshot, at roughly 90% fewer hallucinations than Whisper v2 and roughly 70% fewer than earlier GPT-4o-transcribe models on an internal noise evaluation.^[13] A 90% reduction from a roughly 1% base rate is still not zero, and none of these evaluations are reproducible outside OpenAI.

The current documentation reflects the lesson indirectly. It tells developers to build eval sets containing "names, numbers, dates, alphanumeric strings, and domain terminology," to test with real telephony audio and background noise rather than clean samples, and to "track errors that matter to the application instead of relying only on word error rate," giving medication names and order numbers as examples.^[5] The Realtime checklist adds that empty, truncated, and delayed transcripts should be tracked separately from WER, and that applications need a fallback path for the timestamps, speaker labels, and confidence fields these models do not return.^[15] For a system with no confidence score, an application-level check on the fields it actually cares about is the only available guardrail.

Pricing

OpenAI prices the two new models purely by audio duration. The model cards give the billing metric explicitly as "Transcription audio duration, per minute" for GPT-Transcribe and "Realtime audio duration, per minute" for GPT-Live-Transcribe, with no per-token rate at all.^[3][4] That differs from the previous generation, whose entries still carry per-million-token input and output rates with the per-minute figure shown as an estimate.

Scroll sideways for more →

Source: OpenAI API pricing page, accessed July 31, 2026.^[7]

GPT-Transcribe therefore costs 25% less per minute than gpt-4o-transcribe and than the hosted whisper-1 endpoint, while gpt-4o-mini-transcribe remains cheaper still at $0.003. GPT-Live-Transcribe is priced identically to gpt-realtime-whisper at $0.017 per minute, and half the $0.034 per minute of gpt-realtime-translate.^[7] Streaming costs about 3.8 times as much per minute as file transcription, which is the clearest signal in the pricing that OpenAI expects most volume to be offline.

Applications and integration

The obvious uses follow the split between the two models. GPT-Live-Transcribe suits live captioning, agent-assist overlays in contact centres, real-time meeting notes, and the transcription leg of a chained voice agent, where an application wants explicit control over the transcript before a text agent reasons over it. OpenAI's voice agent guide describes exactly this chained pipeline as the alternative to a speech-to-speech session, recommending it for support flows, approval-heavy flows, and any case where a durable transcript and deterministic logic between stages matter more than the lowest possible latency.^[24] Applications that want a spoken reply generally use a speech-to-speech Realtime model instead, and reach for a transcription session only when text is the product.

GPT-Transcribe suits podcast and video captioning, archive and media transcription, call recording analysis, and clinical or legal documentation, subject to the compliance points below. Because it accepts a prompt, it fits bulk pipelines where each chunk of a long recording can be conditioned on the tail of the previous chunk's transcript, a technique OpenAI lists among the useful prompting scenarios.^[14] The 25 MB request ceiling means long-form media still requires chunking at the application layer.

Third-party voice agent platforms that sit on the Realtime API, including LiveKit Agents, Vapi, and Retell AI, consume transcription sessions through the same interface, as do transcript products such as Otter.ai built on hosted ASR.

Privacy and compliance

Audio is unusually sensitive input: it carries biometric voice data, and in medical or legal settings it carries protected content that the speaker may not know is being processed. Three facts from OpenAI's data-controls documentation are directly relevant.

First, data sent to the OpenAI API has not been used to train or improve OpenAI models since March 1, 2023, unless a customer explicitly opts in.^[25] Second, the retention profile differs by endpoint: /v1/audio/transcriptions is listed with no abuse-monitoring retention and no application-state retention, and is eligible for Zero Data Retention, whereas /v1/realtime carries the default 30-day abuse-monitoring retention and is also ZDR-eligible.^[25] Zero Data Retention and Modified Abuse Monitoring both require prior approval by OpenAI.^[25]

Third, and easy to miss, data residency has not yet caught up with the new models. OpenAI's residency table lists whisper-1, gpt-4o-transcribe, and gpt-4o-mini-transcribe as the supported transcription models on the audio endpoints, and gpt-realtime-whisper as the supported model for /v1/realtime/transcription_sessions.^[25] As of July 31, 2026, neither gpt-transcribe nor gpt-live-transcribe appears on that list, so organizations with a regional storage requirement under GDPR or a comparable regime cannot yet move to them without checking with OpenAI. Processing PHI under a Business Associate Addendum likewise depends on endpoint-level BAA eligibility rather than on the model, and the audio transcription endpoint is marked ineligible for the Eyes Off and Safety Retention controls.^[25]

See also

• Whisper
• Speech recognition
• Automatic speech recognition models
• GPT-Realtime
• Word error rate
• Hallucination
• Voxtral
• Text to speech

References

1. ^OpenAI Developers (@OpenAIDevs). "We're introducing two new transcription models in the API." X, July 28, 2026. x.com/...2082201169443905798
2. ^VeitB. "GPT-Live-Transcribe and GPT-Transcribe: Two New Transcription Models in the API." OpenAI Developer Community, July 29, 2026. community.openai.com/...1388318
3. ^OpenAI. "GPT Transcribe Model." OpenAI API documentation, accessed July 31, 2026. developers.openai.com/...gpt-transcribe
4. ^OpenAI. "GPT Live Transcribe Model." OpenAI API documentation, accessed July 31, 2026. developers.openai.com/...gpt-live-transcribe
5. ^OpenAI. "Transcription: choose the right workflow for recorded or live audio." OpenAI API documentation, accessed July 31, 2026. developers.openai.com/...transcription
6. ^OpenAI. "Changelog." OpenAI API documentation, accessed July 31, 2026. developers.openai.com/...changelog
7. ^OpenAI. "Pricing." OpenAI API documentation, accessed July 31, 2026. developers.openai.com/...pricing
8. ^Radford, Alec; Kim, Jong Wook; Xu, Tao; Brockman, Greg; McLeavey, Christine; Sutskever, Ilya. "Robust Speech Recognition via Large-Scale Weak Supervision." arXiv:2212.04356, December 6, 2022. arxiv.org/...2212.04356
9. ^OpenAI. "openai/whisper" repository metadata (MIT license, 106,255 stars) and "openai/whisper-large-v3" Hugging Face model metadata (5,783,442 monthly downloads), accessed July 31, 2026. github.com/...whisper
10. ^Wiggers, Kyle. "OpenAI upgrades its transcription and voice-generating AI models." TechCrunch, March 20, 2025. techcrunch.com/...n-and-voice-generating-ai-models
11. ^OpenAI. "GPT-4o Transcribe Model." OpenAI API documentation, accessed July 31, 2026. developers.openai.com/...gpt-4o-transcribe
12. ^VeitB. "Introducing GPT-4o Transcribe Diarize: Now Available in the Audio API." OpenAI Developer Community, October 18, 2025. community.openai.com/...1362933
13. ^OpenAI. "Updates for developers building with voice." OpenAI Developers blog, December 2025. developers.openai.com/...updates-audio-models
14. ^OpenAI. "File transcription." OpenAI API documentation, accessed July 31, 2026. developers.openai.com/...speech-to-text
15. ^OpenAI. "Realtime transcription." OpenAI API documentation, accessed July 31, 2026. developers.openai.com/...realtime-transcription
16. ^AlphaSignal. "OpenAI Replaces Whisper With gpt-transcribe, Slashing Errors by 52%." July 28, 2026. alphasignal.ai/...transcribe-slashing-errors-by-52
17. ^Artificial Analysis. "Speech to Text Models: Word Error Rate, Speed and Price Analysis." Accessed July 31, 2026. artificialanalysis.ai/speech-to-text
18. ^The Decoder. "GPT Transcribe improves on its predecessor but can't catch ElevenLabs, Google, or Mistral on error rates." July 29, 2026. the-decoder.com/...oogle-or-mistral-on-error-rates
19. ^Mistral AI. "Voxtral-Small-24B-2507" model card. Hugging Face, accessed July 31, 2026. huggingface.co/...Voxtral-Small-24B-2507
20. ^NVIDIA. "parakeet-tdt-0.6b-v3" model card. Hugging Face, accessed July 31, 2026. huggingface.co/...parakeet-tdt-0.6b-v3
21. ^NVIDIA. "canary-qwen-2.5b" model card. Hugging Face, accessed July 31, 2026. huggingface.co/...canary-qwen-2.5b
22. ^Koenecke, Allison; Choi, Anna Seo Gyeong; Mei, Katelyn X.; Schellmann, Hilke; Sloane, Mona. "Careless Whisper: Speech-to-Text Hallucination Harms." ACM Conference on Fairness, Accountability, and Transparency (FAccT '24), June 2024; arXiv:2402.08021. arxiv.org/...2402.08021
23. ^Burke, Garance; Schellmann, Hilke. "Researchers say an AI-powered transcription tool used in hospitals invents things no one ever said." Associated Press, October 26, 2024. apnews.com/...ess-90020cdf5fa16c79ca2e5b6c4c9bbb14
24. ^OpenAI. "Voice agents." OpenAI API documentation, accessed July 31, 2026. developers.openai.com/...voice-agents
25. ^OpenAI. "Data controls in the OpenAI platform." OpenAI API documentation, accessed July 31, 2026. developers.openai.com/...your-data
26. ^GIGAZINE. "OpenAI releases transcription AI 'GPT Transcribe' and 'GPT Live Transcribe'." July 29, 2026. gigazine.net/...20260729-gpt-live-transcribe