[Bug]: team, end-user, tag, and model budgets admit requests at the exact limit #33321

Check for existing issues

• I have searched the existing issues and checked that my issue is not a duplicate

What happened?

Hard-budget checks use different boundary operators for the same product concept on current litellm_internal_staging at commit 65ca095d4d15a82372e53a547a2390a69f7e1797

Key, user, organization, and centralized team-member checks reject when current spend is equal to the configured maximum by using >=

The following active checks admit at equality:

Relevant code:

• team check
• end-user check
• tag check
• model checks

At the exact hard limit, these scopes allow another request while key, user, organization, and the current centralized team-member path block it

Reservation can prevent some overspend when a reliable maximum request cost is available, but the read-time check remains authoritative when reservation is skipped or cannot estimate the request

The expected behavior is one documented boundary rule for hard budgets, preferably current_spend >= max_budget to match the key, user, organization, and centralized team-member checks

Steps to Reproduce

In focused unit calls, stub the Redis-first spend lookup to return 10.0, configure the relevant maximum as 10.0, and invoke each helper

Equivalent test setup:

current_spend = 10.0
max_budget = 10.0

# Team: _team_max_budget_check with team_object.spend/max_budget == 10
# End user: _check_end_user_budget with spend/max_budget == 10
# Tag: _tag_max_budget_check with tag spend/max_budget == 10
# Model: is_key_within_model_budget with model spend/max_budget == 10

Observed in the current code:

team_equal: allowed
end_user_equal: allowed
tag_equal: allowed
virtual_key_model_equal: allowed
end_user_model_equal: allowed

No BudgetExceededError is raised until spend is strictly greater than the cap

Expected:

team_equal: BudgetExceededError
end_user_equal: BudgetExceededError
tag_equal: BudgetExceededError
virtual_key_model_equal: BudgetExceededError
end_user_model_equal: BudgetExceededError

A regression test should parameterize every hard-budget entity at max - epsilon, max, and max + epsilon

Relevant log output

spend=10.0 max_budget=10.0
actual: request admitted for team, end-user, tag, and per-model checks
expected: hard-budget rejection at equality

What part of LiteLLM is this about?

Proxy

What LiteLLM version are you on ?

Current litellm_internal_staging at 65ca095d4d15a82372e53a547a2390a69f7e1797

Twitter / LinkedIn details

No response