Pricing and cost semantics
How SuperPenguin estimates SDK costs and how that differs from provider-billed amounts.
Estimated SDK cost
When a wrapped call completes, SuperPenguin prices the observed usage against its model catalog (including cache, audio, character, and other meters). Dashboards and alerts use that estimate so you can attribute spend in dollars without waiting for an invoice.
Unknown models are still tracked (tokens, latency, metadata). Estimated cost stays at $0 until pricing exists.
Examples of catalog-backed prefixes and paths:
| Model / path | Provider path |
|---|---|
gpt-*, o3-*, o4-* | OpenAI |
claude-* | Anthropic |
gemini-* | |
grok-* | xAI |
Bedrock modelId | AWS Bedrock (server-side aws_bedrock rates) |
deepgram/nova-* | Deepgram (per audio-second) |
elevenlabs/eleven_* | ElevenLabs (per character) |
Provider-billed cost
Billed cost comes from the provider after you connect an account (or after a provider implements the billing export profile). It reflects invoice semantics: credits, commitments, and settlement state.
For provider implementers, BilledCost must be provider-reported billing data in decimal currency strings. It must not be a public-rate estimate labeled as billed. See Cost data.
What the SDK tracks by default
Each event includes provider, model, usage meters, latency, streaming/tool/vision flags, and attribution metadata. Prompt and completion text are not sent unless content capture is opted in. See Content capture and privacy.
Regional, deployment, service-tier, and batch pricing axes are covered in Batch and service-tier pricing.