gpt-4o-mini-transcribe-2025-03-20

Overview. gpt-4o-mini-transcribe-2025-03-20 is OpenAI's speech-to-text transcription model, released March 20, 2025, as part of the GPT-4o Mini family. It replaces Whisper-based pipelines with a model built specifically for transcription, improving word error rate and multilingual recognition accuracy. Developers reach for it when they need audio converted to text as a discrete, well-defined task rather than as part of a broader chat or reasoning workflow.

Key strengths. OpenAI reports improved word error rates and multilingual accuracy compared to Whisper, the basis this model was built to surpass. It supports streaming of transcript text as deltas, making it suitable for live captioning and voice-agent pipelines where partial results need to appear as audio arrives. As a narrow, single-step audio-to-text task, it fits cleanly into pipelines that don't require multi-step reasoning or tool use.

When to use it. Use this model when the task is transcription itself — live captioning, meeting notes, voice-agent input parsing, or batch conversion of recorded audio to text. Its streaming support makes it a good fit for latency-sensitive, real-time speech-to-text use cases. It is not the right choice for tasks requiring reasoning over the transcribed content, tool calling, or structured output beyond standard transcript formats — those need a separate downstream model.

Limitations. The transcription-specialized variant has a notably smaller context window (16K tokens) and lower max output (2K tokens) than the base gpt-4o-mini instruct model, so it isn't suited to long-audio jobs requiring extensive context retention in a single call. Its training data reflects a cutoff of June 2024. It does not support tool calling, JSON-schema-constrained output, or audio output — transcription only, text out.

Sovereignty note. This model is created by OpenAI, a US company, and inherits US CLOUD Act exposure with no Sovereign or self-hosted deployment path. Audio inputs may carry sensitive or personal content, so teams with strict data-residency requirements should review OpenAI's data handling policies before routing regulated audio through this endpoint.