Source: LiteLLM docs: budgets and rate limits (proxy users, teams, keys)

LiteLLM proxy's budget hierarchy (global, team, team member, internal user, virtual key, end user, per model) and rpm/tpm/parallel-request limits, with reset windows and fail-open/fail-closed behaviour.

Contents
  • What: official documentation page (“Budgets, Rate Limits”)
  • Publisher: BerriAI (LiteLLM)
  • Date: living docs, read 2026-11-14 (mentions behaviour changes in v1.94.0 and v1.95.0)
  • Original: docs.litellm.ai/docs/proxy/users

Summary

The proxy can cap spend (in USD, priced from its model cost map) and request/token rates at several levels of an org hierarchy. Everything except the global budget depends on Postgres, and enforcement details (which budget wins, when windows reset, what happens when the database or Redis is down) are configurable and easy to get wrong.

Key claims

  • Postgres is required. Without it, budgets fail open: requests keep being served.
  • Levels: global (litellm_settings.max_budget + budget_duration), team, team member (max_budget_in_team), internal user, virtual key, end user / customer (via /budget/new and max_end_user_budget_id), and per-model budgets on keys/users (model_max_budget, Enterprise).
  • A key that belongs to a team is checked against team and team-member budgets only, not its owner’s personal budget (briefly enforced in v1.94.0, reverted in v1.95.0).
  • Durations are written like 30s, 30m, 30h, 30d; a key can hold several windows via budget_limits. 24h resets at midnight UTC, 7d on Monday, 30d on the 1st of the month. A background job checks resets about every 10 minutes by default.
  • Over-budget errors include ExceededTokenBudget (key) and ExceededBudget (team member, end user).
  • rpm_limit, tpm_limit and max_parallel_requests can be set on keys, users and teams; model_rpm_limit / model_tpm_limit per model, with key values overriding team values. token_rate_limit_type counts input, output or total (default) tokens.
  • Budget reservation is on by default (it reserves estimated cost before the call); fail_closed_budget_enforcement checks the database before admitting a request; fail_closed_rate_limit_enforcement returns 503 when Redis cannot confirm counters.
  • Agents can carry per-session token/request limits, max_iterations and max_budget_per_session, returning 429 when exceeded.
  • Models priced at zero input and output cost skip budget checks.

Feeds

LiteLLM

Suggest changes →