Short answer: All five major providers now give you a way to watch spend and at least one way to stop it. OpenAI has enforceable hard spend limits at the organization and project level. Anthropic pauses the API when you reach the monthly spend cap you set. Google stops serving when your prepay balance or its billing-account cap is hit. xAI runs on prepaid credits. OpenRouter puts an optional credit limit on every key. Check usage weekly, set alerts at 50% and 80% of your expected month, and set the cap about 30% above your worst real month.

BYOK moves your AI billing to the provider dashboards, which is the point: you pay API rates instead of subscription markup. The catch is that five dashboards is five places a bill can grow unnoticed. A loop that reprocesses a 200K-token file 100 times a day burns 20 million tokens daily: $40 a day on Claude Sonnet 5 input rates, $200 on Claude Fable 5.1.

The controls got better in 2026. OpenAI rolled hard spend limits out to all API accounts in July, Anthropic caps spend by tier and lets you lower it, and prepay balances became the default on Google's Gemini API. Here is what each provider actually enforces, where the settings live, and where the caps still leak.

How to monitor and cap your API spend

Provider	Watch spend	What can stop usage	Where the control lives
OpenAI	Usage dashboard by org and project, plus spend alerts at custom thresholds	Hard spend limit that fails requests (`organization_spend_limit_exceeded` / `project_spend_limit_exceeded`)	Platform settings → Limits
Anthropic	Console Usage and Cost pages (model, workspace, key; CSV export)	Monthly spend cap you set on the Billing page; API pauses until the month resets	Console → Billing
Google	AI Studio Dashboard → Usage; billing status per project	Project spend caps; billing-account tier caps; prepay balance at $0 stops every linked key	AI Studio → Spend / Billing
xAI	Console usage dashboard; per-request cost in every API response	Prepaid credits: with the postpaid limit set to $0, requests stop when the balance empties	Console → Billing; Management API
OpenRouter	Activity page (spend by model and key, CSV/PDF export); key endpoint	Per-key credit limits; an exhausted key returns 402 Payment Required	OpenRouter → keys

Size the cap before you set it

Start from a normal month, not a guess. Using the defaults our calculator ships with (1,000 messages, each with 10K input tokens, 2K output tokens, and 500 reasoning tokens):

Model	Rate per 1M in / out	A 1,000-message month
GPT-5.6 Luna	$0.20 / $1.20	$5.00
GPT-5.4 Mini	$0.75 / $4.50	$18.75
Grok 4.6	$2 / $6	$35.00
Claude Sonnet 5	$2 / $10	$45.00
GPT-5.6 Terra	$2 / $12	$50.00
Gemini 3.1 Pro	$2 / $12	$50.00
GPT-5.6 Sol	$4 / $20	$90.00
Claude Opus 5	$5 / $25	$112.50

Rates read from the providers' live pricing pages on September 17, 2026. If your worst real month on Sonnet 5 was $45, set the cap near $60, with alerts at $30 and $48: room for a heavy week, and two warnings before anything stops.

One lever moves these totals more than any cap does. Prompt caching prices reused input at a steep discount: Sonnet 5 cache hits bill at $0.20 per million against $2 for fresh input, Gemini 3.7 Flash at $0.075 against $0.75. On a 10-million-token month, that is $2 instead of $20 on Sonnet. Here is how cache reads and writes work.

OpenAI: enforce a hard limit in four steps

Organization limits → Spend → Edit spend limit.
Enter the Monthly spend limit.
Turn on Enforce a hard limit.
Select Save.

Organization limits cover every project; project limits cover only traffic billed to that project, and both can apply to a request. At an enforced limit, requests return a 429 with organization_spend_limit_exceeded or project_spend_limit_exceeded. Two caveats: enforcement is not instantaneous, so recorded spend can slightly exceed the cap, and traffic resumes only after a raised limit propagates; otherwise it comes back at the next monthly reset.

Alerts run alongside limits. A project spend limit creates a default alert at the 100% threshold, and you can add your own, so you can get warned at 50% or 80% with the hard limit sitting above as a backstop. OpenAI also assigns each organization an approved monthly usage limit by tier ($100 on Tier 1 up to $200,000 on Tier 5), separate from configured limits. If that one trips, you request an increase rather than edit a setting. Prepaid credits are available too, and they expire after one year.

Anthropic: a spend cap that pauses the API

Each usage tier carries a monthly spend cap: $500 on Start, $1,000 on Build, $200,000 on Scale, with Custom-tier arrangements through an account team. You can set your own lower limit on the Billing page. When the cap is reached, the API pauses with 429s until 00:00 UTC on the first of the next month, unless you raise the limit sooner.

Workspaces add a layer: set custom spend and rate limits per workspace so one project cannot eat the whole budget (the default workspace is the exception). For monitoring, the Console has Usage and Cost pages with breakdowns by model, workspace, and API key, plus CSV export, and the rate-limit charts show your cache rate.

Google: the prepay balance is the real cap

Gemini's API runs on Prepay or Postpay, and new users default to Prepay. You buy credits up front ($5 minimum, $5,000 maximum per purchase), usage deducts in near real time, and when the balance hits $0, every API key in every linked project stops working at once. It is the strictest stop of the five providers, and the most common cause of "why did everything break" for Gemini users.

Auto-reload is optional. If you enable it, set the monthly auto-charge limit that disables reloading for the rest of the cycle once reached; manual top-ups do not count against it.

Project spend caps live in AI Studio under Spend → Monthly spend cap → Edit spend cap. Google's docs flag two caveats: caps are not available on invoiced accounts, and billing data lags about 10 minutes, so brief overages are possible. Long-running work like batch jobs and agent sessions can also overspend project caps. Above them sit billing-account tier caps ($250 on Tier 1, $2,000 on Tier 2, $20,000 to $100,000+ on Tier 3), which pause all linked projects until the first of the month, and spend-based rate limits ($10, $50, or $200 per 10 minutes by tier). Monitoring lives under Dashboard → Usage in AI Studio.

xAI: prepaid by design

xAI bills pay-as-you-go with prepaid credits in the Console, or monthly invoicing for enterprise accounts. The Console has a real-time usage dashboard and a balance you top up on the Billing page. The cleanest way to make it self-limiting: set the postpaid spending limit to $0 via the Management API (POST /v1/billing/teams/{team_id}/postpaid/spending-limits). Prepaid credits are consumed first, and with no postpaid headroom, requests stop when the balance runs out.

Every response also returns the exact cost of that request in cost_in_usd_ticks (1 USD = 10 billion ticks), including streaming and agent loops with server-side tools. Log that field and cost tracking becomes a number your code already has.

OpenRouter: per-key credit limits

OpenRouter is the only one of the five with true per-key hard caps: each key can carry an optional credit limit, plus an include_byok_in_limit switch for whether your own provider keys count against it. That makes it the natural place to fence experiments, teammates, or a script you do not fully trust; an exhausted key returns 402 Payment Required instead of billing more.

To check a key programmatically, GET https://openrouter.ai/api/v1/key returns limit, limit_remaining, and limit_reset alongside usage. The Activity page shows spend (credits plus estimated BYOK usage), tokens, and requests grouped by model, API key, or member, with CSV and PDF export. Purchasing credits carries a 5.5% fee (minimum $0.80; 5% crypto), the overhead you pay for aggregation and control.

Where the caps don't save you

Enforcement is never instant. OpenAI says recorded spend can slightly exceed a hard limit, Google's billing data lags about 10 minutes, and batch or agent jobs can overshoot project caps. Treat caps as circuit breakers, not meters.
Auto-reload can refill what a cap empties. Bound it (Google's auto-charge limit) or set the trigger threshold deliberately (OpenRouter).
Caps cover API usage, not subscriptions. ChatGPT Plus or Claude Pro bills separately and does not count here.
A cap bounds cost; it does not cut it. The levers that cut are caching, model tiering, and reasoning effort (why reasoning tokens bill as output).
Caps reset. A cap that stopped a runaway script stops protecting you when the month flips. Investigate what tripped it before raising it.

Which control should you use?

If you want...	Use this
A hard ceiling on the whole organization	OpenAI org-level hard limit, or Anthropic's spend cap
To box in one script, teammate, or experiment	OpenRouter per-key credit limit
Hands-off budgeting that cannot surprise you	Google Prepay with auto-reload off or bounded
Per-request cost logging your code can read	xAI `cost_in_usd_ticks`
Warnings before anything stops	OpenAI spend alerts at 50% and 80% alongside the hard limit

FAQ

Do AI APIs have hard spending caps in 2026?

Yes, with different mechanics. OpenAI's hard spend limits fail requests once reached, Anthropic pauses the API at your monthly spend cap, Google stops all linked keys at its billing-account cap or when a prepay balance hits $0, and OpenRouter enforces per-key credit limits. xAI's effective cap is its prepaid balance.

What error do you get when a limit is hit?

OpenAI returns a 429 with organization_spend_limit_exceeded or project_spend_limit_exceeded. Anthropic returns a 429 and pauses until the month resets. OpenRouter returns 402 Payment Required. Google returns 429 for spend-based rate limits and stops serving entirely when prepay hits zero.

Do these caps apply to ChatGPT Plus or Claude Pro subscriptions?

No. API spend limits are separate from consumer subscription billing; a ChatGPT Plus month and an OpenAI API month are billed independently.

How do I know what cap to set?

Estimate a normal month from the provider rate card first, add about 30% to your heaviest real month, and set alerts at 50% and 80%. Our cost calculator runs this per model, separating the workspace fee from provider usage.

Can I monitor spend programmatically?

Yes. OpenRouter's key endpoint returns live balance and remaining limits, xAI returns per-request cost in every response, Anthropic exports usage and cost as CSV, and OpenAI's dashboards break usage down by organization and project.

Start your free trial → — 7 days, all providers, no credit card required.

How to Monitor and Cap Your AI API Spending (2026)