How to cap Anthropic API spending
Last verified
Two settings bound your worst case: an organisation spend limit, and automatic credit purchase turned off. Configure both in the Anthropic Console under billing.
The shape is the same as OpenAI's — what differs is the vocabulary and the workspace model.
Where the controls are
In the Anthropic Console, under organisation settings → billing, you'll find spend limits, threshold notifications, and credit configuration.
Anthropic's billing controls have changed as the platform matured, and what's available can depend on your plan. Treat this page as orientation and confirm the exact options in your own console rather than assuming.
The setting that matters most
Automatic credit purchase. Same trap as OpenAI's auto-recharge: it removes the natural cap.
A prepaid balance stops when it's empty, which bounds your loss to whatever was in it. Automatic top-up refills and traffic continues, so a runaway loop or a compromised key can cycle through multiple charges instead of stopping once.
If you turn off exactly one thing, turn off this.
Workspaces as budget boundaries
Anthropic's workspace model is the useful structural tool here. Spend limits apply at organisation and workspace level rather than per key, so workspaces are how you isolate risk.
A sensible layout:
- Production — your real traffic, with a limit sized for normal operation plus headroom
- Development — a low limit, so a local script can't consume the production budget
- Experiments — a very low limit, and treat keys in it as disposable
Keys are scoped to workspaces, which means this also gives you spend attribution for free. It's the same argument as one key per service, one level up.
A configuration that works
Solo developer:
- Organisation limit at what you'd be annoyed to lose — $50–100 is typical
- Threshold notifications at 50% and 90%
- Automatic credit purchase off
- A comfortable prepaid balance so manual top-ups are never urgent
Small team:
- Organisation limit around 3× a normal month
- Separate workspaces with their own lower limits for development and experiments
- Notifications at 50%, 75%, 90%
- Automatic purchase off, or on with a small increment and someone watching
- A daily spend check, because a monthly limit says nothing about a bad Tuesday
What limits don't do
A limit trips at a number. It cannot distinguish a launch from a bug from a stolen key, and by the time it trips the money is gone.
Closing the gap between "something changed" and "the cap engaged" needs a daily number compared against normal. Reading it programmatically needs an admin key; TKN does it alongside your other providers and puts the kill switch on the same screen.