# OpenAI

Models, fine-tuning, batches, files, vector stores, containers and evals, plus organization projects, members, keys and real spend — with a Speech tab for text-to-speech and transcription.

OpenAI's API is really two APIs behind one brand: the project plane your `sk-proj-` key talks to, and the organization plane under `/v1/organization/*` that only an admin key can reach. This plugin covers both, and tells you which half you are missing when you have only supplied one key.

## What you can manage

- **Models** — every model this key can call, base and fine-tuned, each with a **Speech support** panel saying whether it can synthesize, transcribe, or neither. Fine-tuned snapshots can be deleted; base models cannot.
- **Fine-tuning jobs** — create a run from a training file and a base model, then pause, resume or cancel it while it is going.
- **Batches** — create a batch from an uploaded JSONL file, watch the request counters, cancel it. Batches are never deletable.
- **Files** — fine-tuning datasets, batch inputs and outputs, and file-search sources (delete).
- **Vector stores** — the chunked-and-embedded collections behind the `file_search` tool (create, rename, delete).
- **Containers** — the sandboxes Code Interpreter runs inside (create, delete). They expire on their own idle timer and bill per session.
- **Evals** — evaluation definitions with their data source and graders (rename, delete).
- **Projects** — the billing and rate-limit boundary (create, rename, archive). Admin key only.
- **Project API keys** — listed and revoked, never created. Admin key only.
- **Organization members** — role changes between `owner` and `reader`, and removal. Admin key only.
- **Invites** — send and revoke. Admin key only.

## Credentials

OpenAI needs **two keys** to cover its whole surface, and they are not interchangeable — each returns `403` on the other's endpoints.

**API Key** (required) — [platform.openai.com/api-keys](https://platform.openai.com/api-keys). A project key starting `sk-` or `sk-proj-`. This drives models, fine-tuning, batches, files, vector stores, containers, evals and the entire Speech tab.

**Admin API Key** (optional) — [Settings → Organization → Admin keys](https://platform.openai.com/settings/organization/admin-keys). Starts `sk-admin-`, and only an organization owner can mint one. It unlocks:

- **Projects**, **project API keys**, **organization members** and **invites** — without it these four lists are unavailable rather than broken,
- the **usage charts** on models and projects,
- **cost collection**, which is otherwise disabled with a message saying why.

![OpenAI Add-account form showing the required API Key field and the optional Admin API Key field, with the description explaining what the admin key unlocks](https://agent-assets.infrawrench.com/docs-screenshots/plugins/openai/add-account.png)

## The Speech tab

Every model gets a **Speech** tab, with both halves driven by one shared model picker. See [Speech testing](../features/speech-testing.md) for how the panel works generally.

- **Synthesize** posts to `/v1/audio/speech` and returns mp3. Ten voices are offered — Alloy, Ash, Ballad, Coral, Echo, Sage, Shimmer, Verse, Marin and Cedar — which is exactly the current `VoiceIdsShared` enum. `fable`, `onyx` and `nova` still resolve for old integrations but are no longer in the schema, so they are not listed.
- **Transcribe** posts to `/v1/audio/transcriptions`, with `gpt-4o-transcribe`, `gpt-4o-mini-transcribe`, `gpt-transcribe`, `whisper-1` and `gpt-4o-transcribe-diarize` available.

If you open a model that cannot do audio at all and hit Synthesize, the plugin falls back to `gpt-4o-mini-tts` rather than sending a request OpenAI will reject; the same happens in reverse for transcription.

![OpenAI Speech tab on a model, with the voice picker open showing the ten built-in voices and a synthesized clip in the player below](https://agent-assets.infrawrench.com/docs-screenshots/plugins/openai/speech-tab-voice-picker.png)

## Costs and metrics

With an admin key attached, spend comes from `GET /v1/organization/costs`, bucketed by day and grouped by line item, with the project id carried alongside as a tag. Up to a year of history is available, and the most recent three days are re-fetched on each sync because OpenAI restates them.

Models and projects each get a **Metrics** tab charting token usage from `GET /v1/organization/usage/completions`.

## Tips & limits

- **Synthesis is capped at 4,096 characters** per request — the API's own `input` limit — and transcription uploads are capped at 25 MB.
- **Word timings only come from `whisper-1`.** Timestamps require `verbose_json`, and the gpt-4o transcribe family only speaks `json`, `text`, `srt` and `vtt`. Ask any of them for `verbose_json` and you get a 400, so the word table appears for whisper-1 and for the diarize model's segments, and nowhere else.
- **Speaker labels only come from `gpt-4o-transcribe-diarize`.**
- **The legacy `tts-1` and `tts-1-hd` models reject the tone-instructions box.** Sending `instructions` to them is a 400 rather than being ignored, so the plugin drops it for those two.
- **The default TTS model is snapshot-pinned** (`gpt-4o-mini-tts-2025-12-15`) so the same text does not quietly start sounding different between two runs.
- **Project API keys cannot be created through the API.** Only service-account keys can — use the project's **Get credentials** action, which creates a service account in the project and hands you its key. A user-owned project key has to be made in the OpenAI dashboard.
- **Projects are archived, never deleted.** Archived projects still appear in usage and cost reports, which is exactly why they stay visible here.
- **The cost and usage endpoints page differently from everything else.** They ignore the `after` cursor the rest of the API uses and accept 1-day buckets only, 180 at a time.