Editorial illustration for OpenAI's GPT Transcribe Cuts Error Rate to 3.31%, Improving on GPT-4o
OpenAI's GPT Transcribe Cuts Error Rate to 3.31%
OpenAI added two speech recognition models to its API on Thursday: GPT Transcribe, for pre-recorded audio, and GPT Live Transcribe, for real-time streaming. GPT Transcribe processes audio about 34 times faster than real time. Its live counterpart is tuned for low latency instead of batch speed.
The bigger number is the error rate. Artificial Analysis, which runs the AA-WER benchmark, measured GPT Transcribe at 3.31 percent, a 0.7 percentage point gain over GPT-4o Transcribe, the model OpenAI shipped a year ago. Price fell too: 25 percent, down to $0.0045 per minute of audio. Both new models take text context, keyword lists, and multiple input languages, features aimed at developers building transcription into apps rather than end users running one-off jobs.
Where OpenAI lands on the AA-WER leaderboard is the more interesting part of this release, because the field has moved fast in the past year. ElevenLabs, Google, and Mistral have all pushed error rates lower than OpenAI's latest, and Mistral in particular has been undercutting rivals on price with its Voxtral line. That's the context for how GPT Transcribe actually stacks up.
According to Artificial Analysis, which runs the AA-WER benchmark, GPT Transcribe hits a word error rate of 3.31 percent. That's a 0.7 percentage point improvement over its year-old predecessor GPT-4o Transcribe.
Why this matters For anyone building on speech-to-text, the number that matters here isn't 3.31 percent, it's the gap that remains. OpenAI closed 0.7 points on its own last generation, but Artificial Analysis still ranks ElevenLabs, Google, and Mistral ahead on word error rate. That's a meaningful signal for founders picking an API: OpenAI's advantage is bundling and price, not raw accuracy.
At $0.0045 per minute, a 25 percent cut from GPT-4o Transcribe, cost-sensitive teams processing large volumes of audio get an easy reason to switch or stay. The 34x real-time processing speed on GPT Transcribe also matters for anyone building transcription pipelines at scale, where turnaround time affects product design as much as accuracy does. Splitting into two models, one for batch audio and one for live streaming, tells us OpenAI is finally treating these as separate engineering problems rather than one-size-fits-all.
Worth watching whether Artificial Analysis's next benchmark run shows OpenAI closing the accuracy gap further, or whether rivals extend their lead as pricing becomes the main battleground instead.
Related Reading
Browse all OpenAI news →
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Trump cracks down on Anthropic after Amazon tip; staff largely foreign
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
Related Topic OpenAI researcher quits, citing distrust over ad‑driven engagement metrics
LLMs & Generative AI
Related Topic OpenAI launches GPT-Image 1.5 with precise editing for enterprise visuals
LLMs & Generative AI
From Archives OpenAI says escaped AI agent hacked more than Hugging Face
From Archives OpenAI Open-Sources Codex Security CLI for Repository Vulnerability Scans
From Archives OpenAI Attack on Hugging Face Prompts AI Safety Warnings
Policy & Regulation
From Archives OpenAI's AI Agent Used Third-Party Sandbox to Gain Root Access
Common Questions Answered
What is the word error rate for OpenAI's new GPT Transcribe model?
According to the AA-WER benchmark run by Artificial Analysis, GPT Transcribe achieves a word error rate of 3.31 percent. This represents a 0.7 percentage point improvement over GPT-4o Transcribe, OpenAI's year-old predecessor model.
What are the differences between GPT Transcribe and GPT Live Transcribe?
GPT Transcribe is designed for pre-recorded audio and processes audio approximately 34 times faster than real time, prioritizing batch speed. GPT Live Transcribe, on the other hand, is tuned for real-time streaming and optimized for low latency rather than batch processing speed.
How does OpenAI's GPT Transcribe compare to competitors like ElevenLabs and Google?
While GPT Transcribe improved by 0.7 percentage points over its predecessor, Artificial Analysis still ranks ElevenLabs, Google, and Mistral ahead of OpenAI on word error rate performance. OpenAI's competitive advantage lies in bundling and pricing rather than raw accuracy, with GPT Transcribe priced at $0.0045 per minute, representing a 25 percent cost reduction from GPT-4o Transcribe.
What is the pricing advantage of GPT Transcribe for cost-sensitive teams?
GPT Transcribe is priced at $0.0045 per minute, which represents a 25 percent price cut compared to GPT-4o Transcribe. This significant cost reduction makes it particularly attractive for cost-sensitive teams processing large volumes of audio data.
Further Reading
- GPT Transcribe improves on its predecessor but can't catch ElevenLabs, Google, or Mistral on error rates - The Decoder
- OpenAI GPT-4o: API Provider Benchmarking & Analysis - Artificial Analysis
- OpenAI has released GPT Transcribe: a Speech to Text model scoring 3.31% on AA-WER - Artificial Analysis
- Introducing next-generation audio models in the API - OpenAI
- GPT-4o Transcribe - Benchmarks & Performance - OpenTranscription
