GPT-4o Transcribe: model profile

Overview

gpt-4o-transcribe was introduced on March 20, 2025, alongside gpt-4o-mini-transcribe and gpt-4o-mini-tts. It followed OpenAI's earlier speech and audio releases: Whisper in 2022, GPT-4o in May 2024, the Realtime API beta in October 2024, and preview GPT-4o audio models later in 2024. OpenAI positioned the model for developer-facing transcription workloads such as customer call centers, meeting notes, and voice agents, not as a general-purpose ChatGPT model slug.

As of January 13, 2026, OpenAI's changelog states that it currently recommends gpt-4o-mini-transcribe over gpt-4o-transcribe for the best results.

OpenAI says the model builds on GPT-4o, uses specialized audio-centric pretraining data, improved distillation, and an "RL-heavy" training paradigm for speech-to-text. It does not publicly disclose parameter count, layer counts, tokenizer, exact corpus size, exact language coverage in training, or whether the deployed system is fully end-to-end or partly cascaded.

The current model page describes gpt-4o-transcribe as audio-plus-text input with text output, a 16,000-token context window, 2,000 max output tokens, and estimated pricing of $0.006 per minute. The default GPT-4o SKU, by contrast, is a 128,000-context text-and-image model; audio I/O lives in separate GPT-4o audio and realtime variants.

Capabilities and features

• Input: audio plus text; output: text.
• A language hint on the request improves both accuracy and latency.
• A text prompt field can guide style or continue a previous audio segment. The API reference says the prompt should match the audio language.
• The model can return token log probabilities when include=["logprobs"] is requested.
• Output formats: OpenAI's general speech guide says gpt-4o-transcribe and gpt-4o-mini-transcribe support JSON or plain-text outputs, while the typed API reference says those models support only json and reserves text and diarized_json for gpt-4o-transcribe-diarize. The two documentation sources are inconsistent on this point.
• Diarization is available on the related model gpt-4o-transcribe-diarize. With chunking set to auto, the server first normalizes loudness and then uses voice activity detection to choose boundaries. Manual server_vad configuration is supported, along with known-speaker reference clips: up to four speaker references with 2 to 10 second samples.
• Customer-accessible fine-tuning is not publicly documented for gpt-4o-transcribe. GPT-4o proper supports documented fine-tuning, structured outputs, and function calling.

OpenAI family comparison

Language support

OpenAI publishes a supported-language list of 57 languages meeting its support criterion.

The launch post says the model outperforms Whisper v2 and v3 on the multilingual FLEURS benchmark and matches or outperforms other leading models across most major languages.

Language coverage in the training data is not publicly disclosed.

Performance and benchmarks

Vendor-reported: the March 2025 launch post says gpt-4o-transcribe achieves lower word error rate (WER) than original Whisper models, outperforms Whisper v2 and v3 across all language evaluations shown, and on FLEURS matches or outperforms other leading models across most major languages. OpenAI also claims gains in accents, noisy environments, and varying speech speeds. The post does not include a numeric benchmark table for gpt-4o-transcribe.

Vendor-reported: in its December 2025 audio-model update, OpenAI says the gpt-4o-mini-transcribe-2025-12-15 snapshot delivered lower WER than prior models on Common Voice and FLEURS without language hints, and in an internal hallucination-with-noise evaluation produced roughly 90% fewer hallucinations than Whisper v2 and about 70% fewer than previous GPT-4o-transcribe models.

Third-party evaluations:

• The AHELM benchmark paper includes GPT-4o-transcribe family models and reports that gpt-4o-transcribe did not show statistically significant ASR bias conditioned on speaker sex in one analysis, while gpt-4o-mini-transcribe showed a male-speaker advantage.
• The WhisperKit paper describes gpt-4o-transcribe as a frontier cloud baseline and stronger than base GPT-4o for transcription.
• The Qwen3-ASR report compares directly against GPT-4o-transcribe as one of three leading proprietary services.
• The Step-Audio 2 report says it preferred GPT-4o Transcribe over GPT-4o Audio because the former gave stronger results, and calls GPT-4o Transcribe one of the specialized ASR systems with leading-edge performance.
• A 2026 "Back to Basics" ASR paper reports environmental degradation on FLEURS: clipping increased Chinese character error rate (CER) from 6.4 to 17.4, English WER from 2.8 to 8.8, Japanese CER from 3.0 to 7.9, and Korean CER from 4.0 to 10.2.
• The HiKE code-switching evaluation reports GPT-4o-Transcribe as the only LLM-based model in that study to outperform Whisper-Large.
• A phoneme-level study of Ukrainian notes the model tended to output Cyrillic and that the researchers explicitly prompted for Cyrillic transcription, indicating that prompt handling affects measured quality.

Public benchmark evidence

Developer reports on the OpenAI community forum are mixed: some report cases where GPT-4o-transcribe handled background noise or language recognition better than Whisper; others report transcript truncation, odd outputs, prompt leakage in related models, language-enforcement issues, and perceived slowdown or instability that later turned out to be networking-specific.

Comparison with major competitors

Latency and throughput

The gpt-4o-transcribe model page labels speed as "Medium."

A language hint on the transcription request improves latency as well as accuracy.

For low-latency live transcription, OpenAI's lineup uses separate realtime models: the gpt-realtime-whisper page labels speed as "Very fast" and exposes an explicit delay/accuracy tradeoff (minimal, low, medium, high, xhigh) for live transcription sessions. Realtime transcription with gpt-realtime-whisper uses 24 kHz mono PCM, with explicit delay settings and optional turn detection depending on model support. The launch post says developers seeking low-latency speech-to-speech experiences should prefer Realtime API speech models.

Throughput and concurrency limits: Not publicly disclosed.

Deployment and integrations

gpt-4o-transcribe is served through the OpenAI Audio API. The speech-to-text guide says OpenAI historically backed transcription and translation endpoints with whisper-1, and now also supports gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-transcribe-diarize on the transcription side. The guide distinguishes three usage patterns: single-shot file transcription, streaming of completed recordings, and realtime transcription sessions for low-latency live audio.

The model is a developer API SKU, not a ChatGPT model slug. ChatGPT release notes from February 14, 2025 still referred to ChatGPT's voice-to-text dictation feature as Whisper, a month before the API launch of the GPT-4o transcription models. OpenAI's Realtime API announcement framed its speech stack as "similar to ChatGPT's Advanced Voice Mode," indicating that consumer ChatGPT voice and the dedicated transcription API slugs are separate product surfaces.

OpenAI customer stories highlight Retell AI using GPT-4o-based voice systems and Parloa using OpenAI models for enterprise voice-driven customer service. The public stories do not specify whether transcription in those deployments is powered by gpt-4o-transcribe, realtime models, or another routing mix.

Pricing

Development and ownership

gpt-4o-transcribe is developed and operated by OpenAI. For the March 2025 audio-model launch, OpenAI publicly credited four research leads: Christina Kim, Junhua Mao, Yi Shen, and Yu Zhang. Named product leads included Anubha Srivastava, Jackie Shannon, Jeff Harris, Reah Miyara, and Xiaolin Hao. Leadership sponsors included Kevin Weil, Mark Chen, Nick Turley, Olivier Godement, Prafulla Dhariwal, Shengjia Zhao, and Andrew Gibiansky. OpenAI also published broad contributor lists across research, engineering, and product.

Credits from OpenAI's March 2025 4o image-generation release identify Jackie Shannon as ChatGPT Product Lead and Mengchao Zhong and Wayne Chang as ChatGPT Engineering Leads. LinkedIn snippets identify Nick Turley as VP, Head of ChatGPT and Sulman Choudhry as Head of Engineering, ChatGPT. These credits are not transcription-specific.

Several names appear in both the audio launch credits and ChatGPT release credits, including Wayne Chang, Xiaolin Hao, Wanning Jiang, Ola Okelola, and Yilei Qian.

Release history