How do you set a hard spending limit on an AI API?

Alerts, limits and caps are different things

Providers use similar words for different behavior, so check which one you are setting. A spend alert, sometimes called a budget, sends an email or notification and lets traffic continue. A hard spend limit makes the API reject requests once tracked spend reaches the amount. Some providers also impose their own ceiling, such as a usage tier cap, that you cannot raise yourself without a request.

Simon Willison's post that drove today's discussion makes the case plainly: a warning email sent at midnight does not help if a runaway agent keeps spending while you sleep. OpenAI's documentation says it directly: spend alerts do not enforce a cap.

OpenAI: enforce a hard limit per organization or project

OpenAI's spend-limits guide describes two levels. An organization hard limit covers all projects; a project hard limit covers only traffic billed to that project. Both can apply at once.

When a limit is reached, requests return a 429 error with code organization_spend_limit_exceeded or project_spend_limit_exceeded. Traffic resumes when you raise or remove the limit, or at the next monthly cycle. OpenAI also assigns each organization an approved monthly usage limit by tier, which is separate from the limits you configure.

• Organization: Organization limits, Spend, Edit spend limit, enter the monthly amount, turn on Enforce a hard limit, Save.
• Project: Project settings, Limits, Spend, Edit spend limit, enter the amount, turn on Enforce a hard limit, Save.
• Give each app or agent its own project so one runaway job cannot spend the whole organization budget.

Anthropic: a tier cap plus your own lower limit

Anthropic's rate-limits documentation says each Start, Build and Scale tier carries a monthly spend cap of 500, 1,000 and 200,000 US dollars respectively, and that API usage pauses until the next month once the cap is reached unless you request a higher limit. Custom-tier organizations have no monthly cap.

You can set your own spend limit below the tier cap in the Claude Console under Settings, Billing, Spend limits, using Adjust limit or Set limit. Separately, workspaces can carry their own rate limits. Billing and spend limits work differently for Claude Platform on AWS, which bills through AWS Marketplace.

OpenRouter: a credit limit on each key

OpenRouter works on prepaid credits, so the account balance is already a hard ceiling. On top of that, each API key can carry a credit limit, and its management API lets you set limit_reset to daily, weekly or monthly so the key refills on a schedule. A key can also be disabled or set to count your own provider keys (BYOK) toward the limit.

When a key or the balance runs out, requests return a 402 error, and the response metadata names the source, such as openrouter_key_limit or openrouter_credits. You can check limit_remaining at any time with GET /api/v1/key. This makes per-key caps a simple way to give an agent its own budget.

Google Cloud and AWS: caps that pause usage

Google announced Spend Caps on Google Cloud Budgets on July 28, 2026. A cap applies to a single project and service for a monthly period, and once reached Google restricts further cost-incurring usage of that service until you lift it manually. Google lists the Gemini API among supported services and says caps for AI services trigger within minutes. It is a public preview, and fixed commitments such as provisioned throughput keep billing.

AWS spend limits apply to a project and pause it, stopping its resources, when the limit is reached, with notifications at 50, 75 and 90 percent. Data is preserved, but AWS says it permanently deletes a paused project's data if you take no action within 90 days. The feature is part of a new experience that AWS says is being released to a limited number of customers, so you may not have it yet.

Gaps to plan for

• Enforcement lags. OpenAI says recorded spend can slightly exceed the configured amount while the limit propagates, and Google says its AI spend caps trigger within minutes, not instantly.
• A cap is an outage by design. If the capped key serves customers, decide in advance whether you prefer errors or a larger bill, and add lower-threshold alerts so a human sees it first.
• Monthly caps do not stop a fast burn early in the month. Combine them with per-key or per-project limits sized to the job, and with step and token limits inside your agent harness.
• Subscriptions and seat plans for chat apps are separate from API billing; these controls cover API usage.

Limits of this answer

Settings paths, cap amounts and error codes come from each provider's documentation as read on October 4, 2026. Providers change billing controls often, and preview features can change before general availability. These controls were not hands-on tested for this answer.

A useful next action

Pick the largest bill you would accept for a month, set it as a hard limit on the project or key your agent uses, add an alert at half that amount, and trigger the limit once in a test project so you know what error your code will see.