gpt-4o-transcribe

Overview. GPT-4o Transcribe is OpenAI's purpose-built speech-to-text model, built on the GPT-4o architecture and released in March 2025. Unlike the general-purpose gpt-4o chat model, it accepts audio input and returns a text transcript, making it a dedicated choice for teams adding voice input to an existing text-based application. It is served exclusively through Inference on OpenAI's proprietary stack.

Key strengths. OpenAI built this model on audio-centric pretraining data specifically to handle heavy accents, noisy rooms, and varying speech speeds better than the earlier Whisper models. Its single-step audio-to-text task fit means transcription happens directly, without multi-step reasoning or tool use in the loop. Output is consistently returned as syntactically valid JSON, which suits programmatic pipelines that need a predictable transcript format.

When to use it. Reach for this model when you need to convert spoken audio, such as calls or meetings, into text and don't need speaker diarization. It also supports OpenAI's Realtime API for low-latency streaming transcription, though OpenAI now recommends its newer gpt-live-transcribe model for new real-time streaming integrations. It is not a fit for chat, reasoning, coding, or any task requiring text or image input alongside audio.

Limitations. OpenAI has deprecated this model, with a scheduled shutdown date, and explicitly advises against using it for new integrations in favor of gpt-transcribe and gpt-live-transcribe. Its context window (16,000 tokens) and maximum output (2,000 tokens) are considerably smaller than the base gpt-4o model, and it only supports JSON as an output format. Its training data has a mid-2024 cutoff, and OpenAI has not disclosed parameter count, architecture details, or tokenizer specifics for this variant.

Sovereignty note. GPT-4o Transcribe is built by OpenAI, a US company, so data sent to this endpoint carries CLOUD Act exposure rather than Sovereign, EU-only handling. This is a relevant consideration for regulated buyers with strict data residency requirements.