GPT Transcribe error rate improves, but OpenAI trails speech rivals

OpenAI has released GPT Transcribe and GPT Live Transcribe through its API, giving developers a cheaper and more accurate successor to GPT-4o Transcribe. The GPT Transcribe error rate is 3.31% on Artificial Analysis’ AA-WER benchmark, an improvement for OpenAI but still behind rival speech-to-text models from ElevenLabs, Google and Mistral.

GPT Transcribe is aimed at pre-recorded audio files, while GPT Live Transcribe is built for real-time streaming use cases with low latency, according to OpenAI’s API documentation. OpenAI says the file-based model processes audio at about 34 times real time. Both models can take text as transcription context, accept keywords and support multiple input languages.

OpenAI also cut the price. GPT Transcribe is listed at $0.0045 per minute of audio, a 25% reduction from the prior model’s rate, according to the company’s documentation. OpenAI did not disclose usage numbers, customer commitments or revenue expectations tied to the launch.

What is GPT Transcribe's error rate?

Artificial Analysis reports a 3.31% word error rate for GPT Transcribe on its AA-WER benchmark. That is 0.7 percentage points better than GPT-4o Transcribe, the OpenAI model it replaces in this part of the company’s speech recognition lineup.

The result is a measurable gain, but it does not put OpenAI at the top of the benchmark. Artificial Analysis ranks ElevenLabs Scribe v2 first among the cited non-streaming speech-to-text models with a 2.3% error rate. Google’s Gemini 3 Pro follows at 2.9%, and Mistral’s Voxtral Small is listed at 3%.

For teams buying transcription through APIs, the spread is small enough that price, latency, language mix and product integration may matter as much as the headline error number. The benchmark still gives procurement and engineering teams a concrete reason to test alternatives instead of defaulting to OpenAI for speech recognition workloads.

How does GPT Transcribe pricing compare?

OpenAI’s $0.0045-per-minute price puts GPT Transcribe below its predecessor but above at least one recent Mistral offer cited in the same market. Mistral’s Voxtral Transcribe V2 starts at $0.003 per minute, undercutting OpenAI on the published entry price.

The new models also sit beside OpenAI’s broader voice stack. OpenAI’s recently announced Realtime model generation includes GPT-Realtime-Whisper, while GPT Live Transcribe is positioned for streaming transcription. That gives OpenAI multiple speech recognition options inside its API, but the published benchmark data does not show OpenAI winning on raw word error rate.

For founders and platform teams building voice agents, meeting transcription, call analytics or audio ingestion products, the launch is a routine but relevant infrastructure update. OpenAI improved accuracy and lowered cost at the same time. Competitors, based on Artificial Analysis’ ranking and Mistral’s published pricing, still give buyers credible reasons to compare models before standardizing.

This story draws on original reporting from The Decoder.