Ternary
LiteLLM can export proxy spend data to Ternary as FOCUS 1.2 formatted cost reports. This lets you allocate and analyse LLM spend alongside your cloud costs in Ternary's FinOps platform — for example, splitting a shared AI invoice across teams, projects, and API keys.
Overview
| Property | Details |
|---|---|
| Destination | Export LiteLLM usage data to a Ternary AI Gateway cost connection |
| Data format | FOCUS CSV (automatically transformed from LiteLLM spend data) |
| Supported operations | Automatic scheduled export (daily; interval for setup validation) |
| Authentication | Ternary API key + connection ID (per-connection bearer token) |
| Privacy | Cost and usage metadata only — no prompt or completion content leaves the gateway |
Prerequisites
You need three values from Ternary:
- Connection ID (
TERNARY_CONNECTION_ID) — identifies the cost connection the data lands in; it appears in the export URL path. - API Key (
TERNARY_API_KEY) — a per-connection bearer token that authenticates the push. - API host (
TERNARY_BASE_URL) — your Ternary API host, which differs by region (US vs EU).
The Connection ID and API Key are minted when you add an AI Gateway integration in Ternary (Settings → Integrations → Add integration → AI Gateway). The API host is the base URL of the Ternary instance you sign in to — the US and EU regions have different hosts.
Setup
Environment variables
| Variable | Required | Description |
|---|---|---|
TERNARY_API_KEY | Yes | Per-connection bearer token from the Ternary AI Gateway integration |
TERNARY_CONNECTION_ID | Yes | Cost connection identifier (used in the export URL path) |
TERNARY_BASE_URL | Yes | Your Ternary API host (region-specific, US vs EU). No default is assumed. |
TERNARY_EXPORT_FREQUENCY | No | daily (default), or interval for short setup-validation loops. LiteLLM spend is a daily aggregate, so there is no sub-daily cadence. |
TERNARY_EXPORT_INTERVAL_SECONDS | No | Seconds between exports when frequency is interval (intended for short setup-validation / test loops) |
Proxy config
model_list:
- model_name: gpt-4o
litellm_params:
model: openai/gpt-4o
api_key: sk-your-key
litellm_settings:
callbacks: ["ternary"]
export TERNARY_API_KEY="<your-connection-api-key>"
export TERNARY_CONNECTION_ID="<your-connection-id>"
export TERNARY_BASE_URL="<your-ternary-api-host>" # region-specific (US vs EU)
litellm --config /path/to/config.yaml
The proxy registers a background job that exports FOCUS-formatted spend data on the configured schedule.
How it works
On each scheduled run LiteLLM transforms recent spend into FOCUS 1.2 CSV and uploads one file per UTC day in the export window:
- Request an upload URL. For each day with data, LiteLLM calls
POST {TERNARY_BASE_URL}/external-cost-sources/v1/{TERNARY_CONNECTION_ID}/upload-urlwithAuthorization: Bearer {TERNARY_API_KEY}and the body{"day": "YYYY-MM-DD"}. Ternary returns a short-lived signed Google Cloud Storage URL for exactly that day's file, plus the headers the upload must carry. - Upload the day. LiteLLM
PUTs that day's CSV straight to the signed URL with the returned headers. The CSV never passes through Ternary's API.
Each day maps to one file, so re-sending a day replaces it rather than double-counting. If one day's upload fails, the remaining days still upload; the next scheduled run re-sends the window, so a missed day catches up on its own.
Privacy
Only cost and usage metadata is exported — spend, token counts, model, provider, and team/key identifiers. Prompt and completion content never leaves the gateway.
Cost provenance (estimate vs. billed)
Gateway-reported spend is a rate-card estimate, not a billed invoice. Ternary tags every row landed through this connection with its provenance (a cost source of "LiteLLM estimate"), derived server-side from the connection — so estimated gateway spend is never mistaken for, or reallocated as, billed cloud cost. Keep the gateway's spend scoped as the estimate it is to avoid double-counting against a separate provider invoice.
FOCUS field mapping
LiteLLM spend data is transformed into the FOCUS 1.2 schema (the same shared transformer used by the other FOCUS destinations):
| LiteLLM field | FOCUS column | Description |
|---|---|---|
spend | BilledCost, EffectiveCost, ListCost, ContractedCost | Cost of the usage (gateway estimate) |
model | ChargeDescription, ResourceId | Model identifier |
model_group | ServiceName | Model group / deployment |
custom_llm_provider | ProviderName, PublisherName | Provider (openai, anthropic, …) |
api_key | BillingAccountId | Hashed API key |
api_key_alias | BillingAccountName | Human-readable key alias |
team_id | SubAccountId | Team identifier |
team_alias | SubAccountName | Team name |
organization_id / organization_alias | Tags | Organization identifier / display name |
Token breakdown
Per-request token counts — prompt_tokens, completion_tokens, cache_read_input_tokens, cache_creation_input_tokens — which the shared FOCUS transformer does not surface as top-level columns, are carried in the standard FOCUS Tags column as JSON. Ternary reads them from Tags on ingest to enable token-weighted cost allocation. Other metadata (user_id, user_email, model, …) also rides in Tags.
