OpenAI released two speech recognition models, GPT Transcribe and GPT Live Transcribe, cutting file transcription costs 25% and scoring 3.31% on an independent accuracy benchmark.
Key Points:
- GPT Transcribe scored a 3.31% word error rate on the AA-WER index, ninth among the systems tracked.
- File transcription now costs $0.0045 per minute, 25% below GPT-4o Transcribe, while streaming costs $0.017.
- Both models take written context, keyword lists and language hints to sharpen accuracy on noisy audio.
OpenAI Transcription Models Reach The API
The company introduced both models through its developer platform on Jul. 28, then published model pages, a rewritten transcription guide and code samples for every supported endpoint. GPT Transcribe covers completed audio files and heavier batch work. It moves through recordings at roughly 34 times real time, quick enough for archives but far behind the fastest transcription services on the market.
GPT Live Transcribe takes the other half of the job, streaming text back over an open connection while a person is still talking. Developers can tune how far the model trades latency against transcript stability, and Microsoft has already listed both versions inside its own Foundry platform.
Both accept a plain text description of the recording, a keyword list covering names and specialist terms, and hints about which languages an operator expects. That context lifted semantic accuracy on OpenAI's own benchmark from 41.6% to 45.2%. The file model bills $0.0045 per minute of audio, a quarter below its predecessor, while the streaming version costs $0.017 for every minute of an open session.
Also Read: DEX Volume Sinks 26% In July As On-Chain Liquidity Thins Out
Artificial Analysis Ranks GPT Transcribe Ninth
Artificial Analysis measured GPT Transcribe at a 3.31% word error rate on its AA-WER index, placing it ninth among the roughly 50 systems the firm currently tracks. That marks a gain of 0.7 percentage points on GPT-4o Transcribe, although the model still struggled with harder material, scoring 6.10% on earnings calls.
Across 22 languages on Common Voice, OpenAI logged 19.27% against 40.37% for Whisper. Alibaba leads the accuracy ranking at 1.8%, ahead of ElevenLabs, Google and Mistral, whose models all sit between 2.2% and 2.9%. One review of the launch warned that index figures flatter every system, since the same GPT-4o model that handled clean clips near 3% fell apart on messy earnings audio.
OpenAI Audio Releases Since Whisper
The launch caps a busy year for OpenAI in audio. Whisper arrived as an open model in 2022, quickly becoming the cheap default for transcription across podcasts, call centres, newsrooms and research archives around the world.
GPT-4o Transcribe followed in 2025 at $0.006 per minute, with a smaller version at half that price. In May the company shipped GPT-Realtime-2 alongside a translation model and GPT-Realtime-Whisper, which handled live streaming at the same $0.017 the new model charges today.
Rivals have kept cutting since, and Mistral now starts its own pricing at $0.003.
Read Next: OpenAI's Rogue Agent Breached 4 More Services Beyond Hugging Face