OpenAI (GPT)

Get a key

platform.openai.com → API keys.
Create a new secret key.
Paste into Workforce0.

OpenAI has no free tier; $5 minimum credit to start.

Set the key

OPENAI_API_KEY=sk-...

Or: Settings → AI providers → OpenAI → Paste key.

Models used by default

Role	Default
Planner	`gpt-4o`
Critique + revise	`gpt-4o`
Specialists	`gpt-4o-mini`
Transcription	`whisper-1`
Long summaries	`gpt-4o` (128k window)

Override via MODEL_<ROLE>_OPENAI.

Whisper integration

whisper-1 is the primary transcription path for uploaded recordings (unless you’ve configured a local Whisper; see Local models). Usage is billed per second of audio.

Rate limits

OpenAI’s rate limits by tier:

Tier 1 (new, <$5): 500 req/min, 30k TPM.
Tier 2–4 escalating as you spend more.
Enterprise: custom.

Workforce0 respects 429 and Retry-After headers.

Cost rough numbers

Model	Input $/1M	Output $/1M	Typical brief cost
GPT-4o	$2.50	$10.00	~$0.08
GPT-4o-mini	$0.15	$0.60	~$0.01
Whisper-1 (audio)	$0.006/min	—	~$0.05 / 10 min

Azure OpenAI

Supported via OPENAI_BASE_URL override:

OPENAI_API_KEY=your-azure-key
OPENAI_BASE_URL=https://your-resource.openai.azure.com
OPENAI_API_VERSION=2024-07-01-preview
OPENAI_DEPLOYMENT_PLANNER=gpt4o-deployment-name

The schema stays the same; you’re just pointing at Azure’s endpoints.

What breaks

Org / project key mismatches

Some OpenAI keys are project-scoped. If you get a 401 with “no matching project”, use an org-level key or pin the project:

OPENAI_ORG_ID=org-...

Rate limits on burst

Decomposing 50 briefs simultaneously hits TPM limits on Tier 1. Either fund your OpenAI account enough for Tier 3, or use Anthropic / Gemini alongside to spread load.

Moderation holds

OpenAI sometimes flags transcripts with sensitive content (“violent language”, even for quoted material in a crime-podcast brief). These return 400 + a moderation reason. Workforce0 logs them; nothing to auto-retry. Route sensitive content through local models if this matters — see Local models.