Usage & metering

What is counted against your account today, what the caps are, and — just as important — what is not counted. Everything here was measured on 2026-09-04 from the running cluster (the LiteLLM ConfigMap, the pod environment, the Redis store) and from the repository files that deploy it; tests/test_goal_20260904.py fails if the numbers on this page stop matching k8s/litellm/deployment.yaml.

What is metered

Every model call that passes through the LiteLLM gateway — all presets and routes with route: litellm in models.yaml, which is nearly everything in the model picker — is priced from LiteLLM's per-token price table and added to your calendar month total by k8s/litellm/budget_tracker.py, a gateway callback that keeps one Redis key per user and month (budget:mo:YYYY-MM:<user-id>).

Per-user cap	one subscription allowance per calendar month (UTC); it resets on the 1st. Its size is an operator setting (see Where this is defined); you see your position as a fraction, never a dollar figure
Admin accounts	a larger allowance (operator setting)
Max accounts	an operator-assigned, time-limited tier at 20× the standard allowance (k8s/litellm/budget_tiers.json; a tier never lowers an admin's allowance). Ask the assistant for your usage and it reports against the cap that actually applies to you
At the cap	the next request is refused with HTTP 429 — "You have used all of your subscription allowance for this month; requests are refused until it resets on the 1st of the month (UTC)." — until the month rolls over
Attribution	the user id Open WebUI injects server-side (X-OpenWebUI-User-* headers); chat and API-key calls count for the same account, and a sub-agent delegation counts under the user who delegated
Failure mode	fail-open: if Redis is unreachable or the request carries no identity, it goes through unmetered rather than blocked — budgeting is a guardrail, not a gate

Measured 2026-09-04

budget_tracker.proxy_handler_instance is in the live ConfigMap's callbacks; the litellm-proxy pods carry MONTHLY_USER_BUDGET=10.0 and MONTHLY_ADMIN_BUDGET=1000.0; LiteLLM is deployed only by scripts/deploy_litellm.sh (ConfigMap + rollout restart); merges under k8s/litellm/ do not trigger the platform deploy pipeline (#584). Redis holds 11 budget:mo:* keys across 2026-08 and 2026-09 — the tracker is on and accruing. Source: k8s/litellm/budget_tracker.py, k8s/litellm/deployment.yaml, k8s/litellm/config.yaml.

Which price is used

The upstream model's entry in LiteLLM's built-in price map (input + output tokens; images count as their token equivalent). A route whose upstream id the price map does not know bills $0 — the aliases in k8s/litellm/config.yaml exist so that priced ids are used for the cloud models. Local models (Ollama on the fleet, vLLM on MIT Engaging) have no price entry and bill $0 today. (Measured: config comments and the callback's use of response_cost; not re-measured per model.)

A second budget that is not enforced

k8s/litellm/custom_auth.py also stamps a max_budget (its DEFAULT_USER_BUDGET / DEFAULT_ADMIN_BUDGET defaults, DEFAULT_*_BUDGET) on every virtual key. LiteLLM would enforce that figure only with its own database, which this deployment does not run (config.yaml: database_url is off — the Prisma query engine is missing from the image). That number never blocks anything; the Redis tracker above is the enforcement.

What you can see

See your own usage

Ask the assistant "how much have I used this month?" (or "how much of my subscription allowance is left?", "how close am I to the cap?") in a chat whose model has the usage_tools tool enabled. The tool reads your month key — budget:mo:YYYY-MM:<your-user-id> — from the same Redis the tracker writes, and answers in one sentence:

You have used 34% of your subscription allowance for 2026-09; 66% remains. It resets on the 1st of the month (UTC).

At 80% of the allowance the reply adds a warning; when it is used up it says so and explains the 429. It is read-only and takes no arguments: the user is always the signed-in person in the chat (Open WebUI's server-injected __user__), so nobody can look up anyone else's spend. Admins are measured against the admin cap. Two failure replies, both fail-open like the tracker itself (a lookup never blocks a request): if Redis does not answer — connection refused, a DNS stall, a timeout — the reply is "Usage lookup is unavailable right now … try again in a minute"; if the tool itself is misconfigured or broken (a bad Redis URL valve, an unexpected error) it is "Usage lookup is misconfigured or broken (…); please tell an administrator", with the traceback in the pod log, never in the chat. A key whose stored value is not a number is reported as "cannot be determined" naming the key — never as $0.00. Source: usage_tools.py (Open WebUI adapter) over usage_core.py (headless core; the same key/formatting logic is importable without Open WebUI), bundled into one file by scripts/bundle_tool.py; tests/unit/test_usage_tools.py pins the key format to the tracker's and the caps to the Deployment.

(Status 2026-09-04: in the repository, pending registration as a tool in Open WebUI — #525. Until then you learn your position at the 429, or by asking an administrator, who can read your month key in Redis.)

What there is not

No usage page: Open WebUI has none, and LiteLLM's spend UI needs the database that is switched off. The tool above is the self-service surface.

Not metered

• Presets served by a direct Open WebUI connection, which bypass the gateway and its budget: dirk-assistant (Anthropic direct), robot-arm and the other VVUQ specialist models on the quick-tunnel connections, and any model that reaches Open WebUI through a direct connection rather than LiteLLM. (Measured 2026-09-04: seven OpenAI-compatible connections are configured; only the first is the gateway — the others are api.anthropic.com, two trycloudflare quick tunnels, and bridge-mini / glimmer / inkling on kratuvak.com.) When the same id is served by both a direct connection and a gateway alias (glimmer, until the gateway alias is deployed), Open WebUI uses the first connection that lists it.
• Tool-server work: VVUQ Lean compilation, femwell mode solves, GitHub MCP calls, marimo cell execution — their compute is not priced.
• Notebook pods, Knowledge/RAG embedding and retrieval, web search and fetch.
• Sub-agent orchestration overhead beyond the model tokens themselves.

Where this is defined

vocabulary.py (the term users see — subscription allowance) · k8s/litellm/budget_tracker.py (mechanism) · k8s/litellm/deployment.yaml (MONTHLY_* caps, the operator-facing dollar size) · k8s/litellm/config.yaml (callbacks, max_parallel_requests) · models.yaml (route: per preset) · tests/test_goal_20260904.py (pins this page to the deployed values).