Pricing

The SDKs send usage (tokens, characters, searches...). The server turns it into cost from a price catalog that it keeps up to date.

Where prices come from

CostKey syncs three public catalogs every 6 hours: LiteLLM, models.dev and OpenRouter.

• When a call arrives for a model no catalog knows, the server checks the catalogs again within 30 minutes. New models usually appear within hours of launch. Once a price appears, the calls that were waiting are priced.
• Every price version is kept. A call is priced at the rate in force when it happened, and a retired model keeps pricing its history.
• When the catalogs disagree, CostKey uses the price most of them agree on, counting the vendor's own listing as an extra vote. A $0 placeholder never beats a real price.
• A feed that suddenly loses most of its models is ignored, and the last good prices stay.

Which price wins

1. Your project's prices. These are set in Settings → Prices & credits, for negotiated rates or self-hosted models.
2. Rates CostKey maintains for APIs the catalogs miss or list stale: Exa, Firecrawl and Tavily credits, Rumik Silk, ElevenLabs Flash/Turbo, Deepgram Aura, AssemblyAI Universal, Cartesia, Sarvam (converted from INR) and Jina.
3. The synced catalog.
4. A built-in table for older models.

How a call is matched

Call	Priced as
Azure OpenAI or Azure AI Foundry	The Azure price, else OpenAI's list price
Vertex AI	The Vertex price, else the Gemini API price
Bedrock model id, inference profile or ARN	The Bedrock price for that model and region prefix
Dated snapshot (claude-sonnet-5-20260101, gpt-6-sol-2026-09-22)	The exact snapshot, else the model
Fine-tune (ft:gpt-4o-mini-2024-07-18:acme::id)	The fine-tune rate, else the base model
Proxy or gateway with a vendor model name	The vendor's list price

What is billed

• Tokens:
  • input and output
  • cached input at the cache rate
  • cache writes
  • long-context tiers
  • image output tokens
  • Gemini thinking tokens, at the output rate
• Search grounding: per search query on Gemini 3 and later, per grounded request on Gemini 2.x. Google Maps grounding is charged per grounded request.
• Characters for text to speech.
• Audio minutes for speech to text.
• Searches and results for search APIs.
• Credits for scraping APIs.
• Pages.

Free tiers aren't deducted (for example, Gemini's monthly free grounding queries). Batch, priority and flex service tiers are priced at the standard rate. Credit and character prices depend on your plan, so CostKey uses the published pay-as-you-go or mid-tier rate. To use your own plan's rate, set it as a project price.

A model without a price

Settings → Prices & credits lists any model whose calls have no price. Set price offers two ways to fix it:

• Same as a known model. Use this for an Azure deployment or proxy route named something other than the model, such as prod-reasoner priced as gpt-6-sol. The mapping follows catalog price changes.
• Your rates. Enter input and output per 1M tokens, or per 1k characters, per minute, per credit, and so on.

Calls that already arrived are repriced right away. A call without a price is never shown as $0.

Self-hosting

The server ships with a snapshot of the catalog, so a new install has prices before its first sync.

Variable	Effect
COSTKEY_PRICE_SYNC=off	Never fetch the public catalogs. For an air-gapped install, use the shipped snapshot.
COSTKEY_PRICE_CATALOG=off	Turn the catalog off and use only the built-in table and project prices.

To refresh the shipped snapshot:

npm run pricing:snapshot --workspace=@costkey/server