# GPT Transcribe (OpenAI API) - Core.Today AI API

> OpenAI's gpt-transcribe speech recognition, the successor to whisper-1 - plain transcript text with detected languages, optional language hints and keyword spelling. No timestamps. Billed per second of audio (0.174375 credits/s, about 10.5 credits per minute), up to 1 hour / 25MB per request.

- **Provider**: OpenAI
- **Model ID**: openai/gpt-transcribe
- **Category**: Audio & TTS
- **Credits**: 11 per minute of audio (0.174375/s, settled on measured length)
- **Speed**: Fast
- **Quality**: High

## Features
- OpenAI's current transcription model - the replacement for whisper-1 (shut down 2027-02-26)
- Cheaper than whisper-1: about 10.5 credits per audio minute
- Language hints with languages[] (ISO 639-1, regional zh-cn / zh-tw / zh-hk)
- keywords[] to spell names, acronyms and domain terms correctly
- Accepts mp3, mp4, mpeg, mpga, m4a, wav and webm up to 25MB

## Use Cases
- Meeting, lecture and interview transcripts
- Voice memos and call recordings to searchable text
- Moving a whisper-1 integration off the retiring model

## API Endpoint

Base URL: https://api.core.today/v1
Create Prediction: POST /predictions
Get Status: GET /predictions/{job_id}

## Authentication
Header: X-API-Key: YOUR_API_KEY

## Input Parameters

### Required
- **audio**: string - Audio file URL (mp3, mp4, mpeg, mpga, m4a, wav, webm). Sent to OpenAI as the `file` part. Max 25MB and 1 hour.

### Optional
- **response_format**: string - json only - {text, languages, usage}. verbose_json, text, srt and vtt are not available for this model (no timestamps).
  Options: json
- **languages**: array - Language codes spoken in the audio (ISO 639-1 such as ko, en, ja; regional zh-cn / zh-tw / zh-hk). Omit for automatic detection.
- **keywords**: array - Names, acronyms and domain terms to spell correctly, one per entry (no line breaks or angle brackets).
- **prompt**: string - Free-form context about the audio (topic, speakers, style).

## Examples

### Transcribe with a language hint
Returns {text, languages, usage} - usage.seconds is the billed length.
```json
{
  "model": "openai/gpt-transcribe",
  "input": {
    "audio": "https://replicate.delivery/mgxm/e5159b1b-508a-4be4-b892-e1eb47850bdc/OSR_uk_000_0050_8k.wav",
    "languages": [
      "en"
    ]
  }
}
```

### Korean meeting with keywords
keywords keeps product and company names spelled the way you write them.
```json
{
  "model": "openai/gpt-transcribe",
  "input": {
    "audio": "https://example.com/meeting.m4a",
    "languages": [
      "ko"
    ],
    "keywords": [
      "코어닷투데이",
      "Core.Today"
    ]
  }
}
```

## Response Format
```json
{
  "job_id": "abc123",
  "status": "pending | processing | completed | failed",
  "result": "URL or data (when completed)"
}
```

## Usage Flow
1. POST /predictions with model and input -> receive job_id
2. GET /predictions/{job_id} -> poll until status is completed or failed
3. Result contains output URL(s)

## Tips
- The result is OpenAI's response body as-is: text, the detected languages, and usage.seconds (the billed length)
- Need segment or word timestamps? gpt-transcribe returns none - use openai/whisper (Replicate, segments with start/end), or openai/whisper-1 until it shuts down on 2027-02-26
- Split files over 25MB or 1 hour before submitting - longer or larger files are rejected
- OpenAI SDK: point the client at base_url https://api.core.today/llm/openai/v1 with your cdt_ API key and call client.audio.transcriptions.create(model="gpt-transcribe", file=...) - the file is uploaded directly (no URL), json only (no streaming, no translations), billed per second at the LLM rate (0.13935 credits/s, about 8.4 per minute) on the length OpenAI reports

## Documentation
https://developers.openai.com/api/docs/models/gpt-transcribe