AI spending limits compared: what each provider lets you cap.

01What changed in 2026.

Until this year, most AI spending controls were alerts. An email arrived when spend crossed a threshold, and usage carried on. In 2026 the providers added limits that stop.

GitHub moved Copilot to pooled AI credits on 1 June, with budgets that can stop usage. OpenAI added hard spend limits for organisations and projects on 22 July. Google put spend caps on Cloud budgets into preview on 29 July. On 16 September GitHub made budget increase requests generally available, so a blocked member can ask for more and an owner can approve it.

TakeawayIn 2026 the providers moved from alerts to limits that stop, and two of them added a way to request more.

02The native controls, side by side.

The table shows what each provider lets you cap, what happens when spend reaches the limit, and whether there is a route to ask for more. It was checked against each provider's own documentation on 21 October 2026.

TakeawayThe model APIs cap accounts, the seat products cap people, and the clouds mostly alert.

03How to set a spending limit, provider by provider.

OpenAI API.

In the platform settings, open Limits for the organisation or for a project and set a monthly budget. Choose an alert, or enforce a hard limit, which makes new requests fail with an error once spend passes it. Enforcement takes a moment to catch up, so spend can run slightly over. OpenAI also sets its own monthly ceiling for each organisation, based on its usage tier.

Anthropic API.

Each usage tier carries a monthly cap, and an organisation can set a lower limit of its own. In the Claude Console, set a monthly spend limit and alert thresholds on each workspace. The default workspace takes no spend limit, so put production work in workspaces of its own. Per-user monthly limits are available in the Claude Code workspace. At a limit, requests stop until the limit is raised or the month resets.

Claude Team and Enterprise.

Owners set an organisation-wide monthly limit, then limits by seat tier, group and member. A member's own limit overrides the group's, and the organisation limit is the ceiling. With no limit set anywhere, spend is uncapped. Members who reach their limit can request more credits, and admins approve.

Claude Code.

Claude Code takes the limits of the account it runs on: member limits on Team and Enterprise, the Claude Code workspace in the Console, or the cloud's budget controls when it runs through Amazon Bedrock, Google Cloud or Microsoft Foundry. Anthropic puts the average cost at around $13 per developer per active day, or $150 to $250 a month.

ChatGPT Enterprise.

Admins set a monthly limit per user, as a workspace default, a group default or an override for one person, and a budget or overage limit for the workspace. By default the workspace has no overage limit. Members who reach their limit can request more from inside ChatGPT, and admins approve a temporary or permanent increase. Alerts notify and leave usage running.

Google Gemini API and Google Cloud.

Gemini API billing tiers carry monthly caps that pause every linked project when reached. In AI Studio, project spend caps are available as an experimental feature. In Google Cloud, spend cap budgets, in preview since July 2026, pause new usage of the Gemini API and the agent platform for one project and one service once estimated spend passes the budget. Standard Cloud budgets only send alerts.

Azure OpenAI.

Microsoft's documentation says Azure OpenAI has no native hard spending limit. Cost Management budgets on a subscription or resource group send alerts, and stopping usage needs an action group connected to automation you build. Cost data arrives 8 to 24 hours late. Quotas limit tokens per minute, which slows usage and leaves the dollar total open.

Amazon Bedrock.

Bedrock has no spend cap of its own. Create an AWS budget for Bedrock cost and attach a budget action that applies an access policy denying model calls, set to run automatically or after manual approval. Budget data updates up to three times a day, so spend can pass the threshold before the action runs. Application inference profiles with cost allocation tags show spend by application.

Cursor.

Admins set a monthly spend limit for the team. Enterprise adds limits by member and by group, and the highest applicable limit applies. At the limit, paid usage stops for the affected users. Cursor notes that enforcement is not instant.

GitHub Copilot.

Since 1 June 2026 Copilot is billed in AI credits pooled across the enterprise. Enterprise owners and billing managers set budgets for the enterprise, cost centres and users, and organisation owners set organisation budgets. User budgets always stop at the limit. Other budgets stop only when "Stop usage when budget limit is reached" is switched on, which it is not by default. Blocked members can request an increase.

TakeawayTurn on the hard limit wherever one exists, and know which of your providers can only send an alert.

04Where native limits stop.

• One provider each. Nothing native adds OpenAI, Anthropic, Google, Azure, AWS, Cursor and GitHub into one number or one limit.
• Caps on accounts. The model APIs limit organisations, projects and workspaces. None sets a limit for a single agent, so one agent on a shared project can spend the project's month.
• Alerts on the clouds. Azure has no dollar cap, AWS needs a policy action, and Google's cap is a preview for one project and one service.
• Late stops. Nearly every provider says enforcement lags and spend can run over before the limit holds.
• Approval in the seat products only. On the model APIs, whoever holds the admin role raises the limit, with no request, reason or expiry on record.

TakeawayNative caps stop a runaway account. They leave the total, the agent and the approver to you.

05How to set one budget across every provider.

1. List every account. Every provider, organisation, project, workspace and seat plan in use, with an owner for each.
2. Turn on native caps as a backstop. Set hard limits wherever they exist, a little above planned spend, so a runaway agent stops even if everything else fails.
3. Set one budget above them. An organisation ceiling, an allocation per team, and caps per agent and per person, counted across every provider before each call is paid for.
4. Route exceptions to one approver. One request route for every provider, with a named approver, a reason and an expiry on each increase.
5. Check the limits monthly. The providers changed these controls four times this year. Re-read their limits pages each month and adjust the backstops.

Our guide to token budgets and spending limits for AI agents covers what should happen at each cap, and the essay on corporate AI spending limits covers how Uber, Walmart, Meta and Tesla set theirs.

TakeawayList the accounts, turn on native caps as a backstop, set one budget above them, route exceptions to one approver, and check monthly.