How do I set a hard spend limit on LLM APIs?

Set it at the provider and per key or project. As of November 4, 2026: OpenAI — Organization or Project limits, Edit spend limit, turn on Enforce a hard limit (requests then return 429 organization_spend_limit_exceeded or project_spend_limit_exceeded). Anthropic — Claude Console Billing, Spend limits, Set limit (requests return 400 invalid_request_error), plus per-workspace limits. OpenRouter — a credit limit on each API key with an optional daily, weekly or monthly reset. Vercel AI Gateway — budgets per team, project, API key or member.

Do LLM spend limits stop spending instantly?

No. OpenAI says enforcement is not instantaneous and recorded spend can slightly exceed the configured amount. Anthropic says Message Batches may go slightly over a workspace spend limit because of concurrent processing. Set the hard cap below the amount you can actually afford, and add an alert at 50 to 80 percent.

What is the default Anthropic API monthly spend cap?

Each Anthropic usage tier has a monthly spend cap: Start $500, Build $1,000 and Scale $200,000; the Custom tier has none. When an organization reaches its tier cap, API usage pauses until 00:00 UTC on the first day of the next month and requests return HTTP 429 with error_code enforced_spend_limit_reached and no retry-after header. These are the figures in Anthropic's rate-limit docs on November 4, 2026.

Does a Baseten budget stop my GPU deployments?

No. Baseten's Enforce budget option rejects only Model API requests once spend reaches the budget. It never stops dedicated deployments or training jobs, which keep running and billing. Cap dedicated spend with max_replica on each deployment.

Quick Answer

How to Set Hard Spend Limits on LLM APIs (2026 Guide)

Published: November 4, 2026

The short answer

Put a hard cap at the provider, a smaller cap on each key or project, and an alert well before either. Most bill shocks come from one leaked key or one runaway agent loop, so the per-key cap matters more than the account cap. Where to set them, as of November 4, 2026:

Platform	Hard cap at	What happens at the cap
OpenAI	Organization and each project	`429` `organization_spend_limit_exceeded` / `project_spend_limit_exceeded`
Anthropic	Organization and each workspace (not the default workspace)	`400` `invalid_request_error`, “You have reached your specified … API usage limits”
OpenRouter	Each API key (credit limit, daily/weekly/monthly reset)	Key stops working until reset or raise
Vercel AI Gateway	Team, project, API key or team member budgets	Requests blocked by the budget
Baseten	Workspace budget with Enforce budget	Only Model API requests rejected; GPU deployments keep running

Taken from each vendor’s documentation on November 4, 2026. For pricing strategy around these costs, see how to price an AI product around token costs.

Step 1: Split spend by project, key or workspace

One key per app, environment and customer-facing agent. A cap on the shared org key alone means one bad loop takes production down with it.

OpenAI: create a project per environment; each project gets its own rate and spend limits and a service-account key.
Anthropic: create a workspace per app; you can set custom spend and rate limits per workspace, but not on the default workspace.
OpenRouter: create keys through the Management API with a limit and limit_reset (daily, weekly or monthly); workspace-scoped management keys can only manage keys in their own workspace.

Step 2: Set the hard caps

OpenAI. Go to Organization limits (or a project’s Settings → Limits) → Spend → Edit spend limit, enter the monthly amount, turn on Enforce a hard limit, save. An org limit applies across every project; a project limit only to that project. Traffic resumes after you raise the limit or at the next monthly cycle.

Anthropic. In the Claude Console, Billing → Spend limits → Set limit. Your own limit cannot exceed your tier’s cap: Start $500, Build $1,000, Scale $200,000 a month (Custom has no cap). Hitting your own limit returns 400 invalid_request_error; hitting the tier cap returns 429 with enforced_spend_limit_reached and no retry-after header, and pauses usage until 00:00 UTC on the 1st. Claude Code workspace limits are checked separately and can return a 429 with retry-after.

OpenRouter. Give every key a credit limit; use limit_reset: "daily" for agents so a runaway loop burns one day’s budget, not a month’s. include_byok_in_limit controls whether your own-key usage counts.

Vercel AI Gateway. Set budgets per team, project, API key or member. Note that exhausting a provider grant does not necessarily stop requests if another balance or automatic recharge can fund them — the budget is the control.

Baseten. Set a monthly budget (emails at 75%, 90%, 100%) and tick Enforce budget — but it only rejects Model API requests. For dedicated GPU deployments the real ceiling is max_replica; see Baseten billing and cost control.

Step 3: Add alerts below the cap

OpenAI spend alerts and Baseten budget emails keep firing alongside hard limits. Put the first alert at 50% and a second at 80% so someone can react before traffic stops.

Step 4: Handle the cap in code

A capped request is not a rate limit. Retrying it — including your SDK’s automatic retries — fails until the limit changes.

except APIStatusError as e:
    code = getattr(e, "code", None) or (e.body or {}).get("error", {}).get("code")
    if code in ("project_spend_limit_exceeded", "organization_spend_limit_exceeded",
                "enforced_spend_limit_reached"):
        disable_feature_and_page_oncall()   # do not retry
    elif e.status_code == 429:
        backoff_and_retry()

Match on the error code, not only the status: OpenAI’s spend cap and its rate limit both return 429, and Anthropic’s own-limit response is a 400.

Step 5: Plan for overshoot

Caps lag. OpenAI states enforcement is not instantaneous and spend can slightly exceed the limit; Anthropic warns batches may go slightly over a workspace limit. Set the hard cap at the amount you would be annoyed to pay, not the amount that would hurt.

Common mistakes

One key everywhere. A leak or loop has no blast radius limit.
Alert-only limits. Alerts do not stop traffic; only “enforce” settings do.
Forgetting GPU spend. Token caps do not cover rented GPUs — cap replicas.
Retrying a spend-cap error. It wastes calls and hides the outage from your on-call.

Last verified: November 4, 2026.