Budgets

parry.install(max_cost_usd=10.00)

Once this guard has spent $10, every call that isn’t already cached raises parry.BudgetExceeded before the request leaves your process. Cache hits are free, so they are always served.

try:
    client.chat.completions.create(model="gpt-4o", messages=msgs)
except parry.BudgetExceeded as e:
    print(f"stopped at ${e.spent_usd:.4f} of ${e.limit_usd:.2f}")
    print(f"next call to {e.model} was estimated at ${e.estimate_usd:.6f}")

How the check works

1. Before each uncached call, parry reads the session spend from the ledger.
2. It estimates the prompt’s input cost: a tiktoken count of the prompt text, priced at the model’s input rate.
3. If spent >= max_cost_usd or spent + estimate > max_cost_usd, it raises BudgetExceeded and counts it in stats.budget_rejections.
4. Otherwise the call goes ahead. Its real cost, output included, is recorded when the response arrives.

The call that crosses the line, usually because of its output, is allowed to finish, since it has already been paid for. Every call after it is refused.

Guarantees and limits

Guarantee	Detail
No call starts after the limit is reached	checked before rate limiting and before the network
The overshoot is at most one call per thread	concurrent threads can each have one call in flight when the line is crossed
Cache hits are always served	they cost nothing
Unpriced models are invisible to the budget	add a price with guard.set_pricing(...)
The scope is the guard’s session	guard.reset_stats() restarts the budget

Budgets and persistence

max_cost_usd covers the current guard session, not lifetime spend stored in SQLite. For a daily cap, compare stats.today_cost_usd yourself, or create a fresh guard each day.

Changing the budget at runtime

guard = parry.get_guard()
guard.max_cost_usd = 25.0     # raise it
guard.max_cost_usd = None     # remove it

Using the budget without the hooks

guard.check_budget("gpt-4o", prompt_tokens=parry.count_tokens(prompt, "gpt-4o"))

The LangChain and LlamaIndex handlers call this before every LLM event.