Pricing database

Every cost number Margint shows is computed the same way:

cost_microdollars = (input_tokens * input_cost_per_million
                  + output_tokens * output_cost_per_million) / 1,000,000

The interesting part is where *_cost_per_million comes from.

Source of truth: LiteLLM

We mirror LiteLLM's canonical pricing registry — the same MIT-licensed file used by 400+ open-source projects and proxy gateways. It covers 400+ models across OpenAI, Anthropic, Google, Mistral, Cohere, Bedrock, Azure, Groq, Fireworks, Perplexity, together.ai, and more.

LiteLLM updates it within hours of provider price announcements. We pull it on a schedule and publish snapshots at /docs/concepts/pricing-database per release.

Local calculation

Both SDKs bundle a snapshot of the pricing database at install time. When you call track() without costMicrodollars, cost is calculated locally — no network hop. This means:

• Your ingestion endpoint never blocks on a pricing lookup.
• Offline / airgapped deployments still get accurate cost numbers.
• SDK-only "local mode" works without a Margint account (handy for development).

Keeping pricing fresh

• SDKs ship with pricing at build time. Reinstall to pick up new models or price changes.
• The server uses the latest database on every ingest. If you pass an unknown model, we record token counts with cost_microdollars = 0 and flag it in the dashboard.
• Override locally by passing costMicrodollars directly on track().

Requesting a model

We're happy to add models. Email hi@margint.dev with the provider name and a link to their pricing page.

Most new frontier models are priced by LiteLLM within 48 hours of release — usually we don't have to do anything.