Skip to content

chatgpt-subscription.md

Latest commit

 

History

History
40 lines (27 loc) · 4.43 KB

chatgpt-subscription.md

File metadata and controls

40 lines (27 loc) · 4.43 KB
Raw

Use your ChatGPT subscription (beta)

Cutawan can use Codex CLI signed in with ChatGPT for clip selection, visual analysis and captions. It uses your account's Codex allowance. A ChatGPT subscription is not an API key, and access/model availability depend on your plan. This option never falls back to a paid API or another model.

Audio transcription runs locally with faster-whisper. Rendering and speaker tracking also run locally. Transcript text and selected frames go to Codex for analysis. Cutawan does not read, copy or store Codex authentication tokens.

One-time setup

  1. Install the current Codex CLI, run codex login, and choose ChatGPT sign-in. codex login status should report ChatGPT, not API-key authentication.

  2. Install Python 3.10+. On first launch, choose ChatGPT sign-in in the setup wizard, enter your Python executable if it is not on PATH, and click Install local Whisper. Cutawan creates a private environment in its app-data folder, installs faster-whisper, and downloads your chosen model from Hugging Face only after you click the button. Choose Small for a lighter download or Large v3 for more accuracy and a multi-GB download. The installer can be cancelled and retried.

    If you prefer to set up Python and the model yourself, create a dedicated environment:

    python -m venv cutawan-whisper

    On Windows, install using cutawan-whisper\Scripts\python.exe -m pip install faster-whisper huggingface-hub. On macOS/Linux, use cutawan-whisper/bin/python -m pip install faster-whisper huggingface-hub.

  3. For manual setup, explicitly download a compatible model. For the large-v3 model used in our quality experiments, run this with that environment's Python (requires several GB of download and storage):

    python -c "from huggingface_hub import snapshot_download; snapshot_download('Systran/faster-whisper-large-v3', local_dir='whisper-large-v3')"

    CPU transcription works but can be slow. A compatible NVIDIA CUDA installation enables faster GPU transcription; see faster-whisper's installation requirements. Cutawan does not install Python or GPU libraries. A smaller compatible model is possible, with a transcription-quality tradeoff.

  4. The first-run wizard saves the installed Python and model paths for you. For manual setup, in Settings → General → AI connection, choose ChatGPT subscription via Codex (beta). Set the Python executable to the environment's full executable path and the speech model folder to the downloaded folder containing model.bin. If Codex is not on PATH, set its executable path too.

  5. Choose Save and check setup. This checks CLI authentication, Python dependencies and model files without making a model request. The first real transcription verifies runtime/GPU compatibility.

Control usage

  • The analysis default is gpt-5.6-luna, low reasoning. A model unavailable to your account produces an error. You can enter another supported model explicitly.
  • The default limit is 10 new requests per UTC day, shared across Cutawan projects on this profile. One video can require more than 10 requests. Set a higher cap deliberately if needed, or set 0 to use only cached analysis.
  • Successful identical requests are cached locally, keyed by model, reasoning setting, messages and output schema. A model change does not reuse another model's cache.
  • Failed inference attempts also count. Cutawan does not automatically retry subscription inference. The limit is a request count, not a token or monetary budget; long prompts and images can consume more allowance.
  • Your account's own credits, subscription limits and spending settings still apply. Cutawan does not buy credits or redeem resets.
  • Cached responses are stored in subscription-cache under the app's data folder and can contain derived information about your media. Temporary prompts/images are removed after each request.

The API connection remains available separately. Existing installations keep their current API settings until they explicitly switch providers. Existing project files need no conversion.

This is an optional CLI integration, not an OpenAI-sponsored product or unlimited ChatGPT API access. It uses the documented non-interactive CLI and ChatGPT authentication paths.