How to pay for AI APIs and agent tools with one card

Why does the card on file matter so much for AI bills?

Most AI vendors bill from prepaid credit, and the card on file is what refills it. If a refill fails and the balance runs out before anyone adds more, your API requests start failing.

OpenAI, Anthropic, xAI, Cursor Teams and Google’s Gemini Prepay plan all run on a credit balance. Each lets you turn on automatic top-ups, which charge the card on file when the balance drops below a number you pick. That works until a charge declines. Then nothing is topping up the balance, it keeps falling, and once it runs out the API starts returning billing errors at whatever hour that happens to be.

A lot of what’s written about cards for AI agents is about something else, cards an agent uses to buy things on its own. Slash makes this point in its piece on cards for AI agents. Paying OpenAI, Anthropic and other AI vendors from a card is, in its words, “a different job.” That job is what this guide covers. We went through each vendor’s own billing docs on November 5, 2026, and the details below come from them.

What happens when an auto-reload charge declines?

That depends on the vendor, and not all of them tell you.

OpenAI tells you the most. Its prepaid billing article says a failed recharge sends you an email, and that on prepaid billing API usage stops once the available balance runs out. From then on requests come back with a billing error, code credit_balance_exhausted. Access might not stop the moment you hit zero, and anything used in that gap becomes a negative balance taken out of your next purchase. One line in there is easy to miss. Manual top-ups don’t count toward the monthly recharge limit, so a manual purchase going through tells you nothing about whether auto-recharge works.

Anthropic is short on this. Its billing article says that once your credits run out you can’t call the API or use the playground until you add more. It doesn’t say what happens when an auto-reload charge fails.

Google’s rule is the harshest we found. On the Gemini Prepay plan, an empty balance stops every API key in every project linked to that billing account at the same moment, and requests fail with HTTP 402 until you add credits. There’s a second catch in Google’s Gemini billing docs. If the same billing account also pays for other Google Cloud services on Postpay, missing a payment there can suspend Gemini API access, however many prepaid credits you have left.

xAI’s docs now go by SpaceXSI, but the rule is plain. Requests get rejected automatically once prepaid credits run out, unless you have invoiced billing, which is off by default. The billing docs suggest turning on auto top-up so you don’t get cut off.

Firecrawl returns HTTP 402 when credits run out and pay-as-you-go is off. Turn pay-as-you-go on and it adds $5 increments, charging the card on file each time the balance reaches zero, up to a monthly cap you set (billing docs).

So on prepaid billing, a declined top-up usually doesn’t break anything right away, because service keeps running on whatever credit is left. Google’s Postpay rule is the one to watch, since a missed payment there can suspend Gemini even with credit left.

How does auto-reload work at each vendor?

They all refill a prepaid balance from a card. Here’s what each one documents.

Vendor	How usage is paid	Auto-reload	Your own cap	When the balance runs out
OpenAI	Prepaid credits for new accounts. $5 minimum purchase, $10 default	On by default at setup. $5 minimum recharge, optional monthly limit	Monthly recharge limit, plus spend limits	Billing error, credit_balance_exhausted
Anthropic	Prepaid credits. Monthly invoicing through Sales	Trigger balance and target balance. Charges right away if you’re already below the trigger	Spend limit below your tier’s cap	No API calls until you add credits
Google Gemini	Prepay or Postpay on a Cloud Billing account	Optional, with a monthly auto-charge limit	Project spend caps in AI Studio (experimental)	Prepay: every key on the billing account stops, HTTP 402
xAI	Prepaid credits. Invoiced billing for Enterprise	Threshold top-up, $5 minimum, monthly maximum	Invoiced billing limit and top-up cap	Requests rejected
Cursor Teams	Seats, plus a shared prepaid balance for usage beyond each seat’s pool	Auto Top-up below a threshold	One monthly additional usage limit for the team	Paid additional usage pauses
Firecrawl	Plan credits each month, plus optional pay-as-you-go	$5 increments charged to the card on file	Monthly pay-as-you-go cap	HTTP 402
OpenRouter	Prepaid credits. Invoicing on Enterprise	Auto top-up	Not covered on the pricing page	Not covered on the pricing page

Sources for each row: OpenAI, Anthropic, Google, xAI, Cursor, Firecrawl and OpenRouter.

Some of these rules trip people up. Cursor’s team limit pauses paid additional usage until next month, or until an admin raises it, even if prepaid credit is still there. At OpenAI and Google, reaching a monthly reload limit stops automatic purchases until the next month, which from the outside looks a lot like a decline. OpenRouter charges a fee when you buy credits, 5.5% on its Standard plan and 8% on Business, and its FAQ says it doesn’t mark up inference prices.

Do usage tiers depend on how you pay?

At the three big model vendors, yes. Tiers go up as you pay more or build usage history, and a higher tier means a higher monthly spend ceiling.

Vendor	How you move up	Spend ceiling
OpenAI	Total credit purchases of $5, $100 and $500 reach Build, Launch and Grow	$500, $5,000 and $200,000 a month for those tiers
Anthropic	Usage history and account standing	Monthly spend caps of $500, $1,000 and $200,000 for Start, Build and Scale
Google Gemini	Tier 2 at $100 paid plus 3 days since the first successful payment. Tier 3 at $1,000 paid plus 30 days	Billing tier caps of $2,000 for Tier 2, and $20,000 to $100,000 or more for Tier 3

Figures come from OpenAI’s rate limits guide, Anthropic’s rate limits page and Google’s Gemini billing docs. If you hit the tier cap at Anthropic, API usage pauses until 00:00 UTC on the first day of the next month, unless you get a higher limit sooner.

Google counts tiers, rate limits and caps at the billing account level, across every project linked to it. That’s why we’d change the payment method on the billing account you already have when you switch cards. Linking your projects to a new billing account means earning that payment history all over again.

How do teams keep one card on file across vendors?

Mostly by going vendor by vendor through the same short routine.

1. To see what the card will carry, list each vendor that charges it, how that vendor bills (prepaid top-ups, metered usage or seats) and who holds the billing role there. At Anthropic only an Admin or Billing role can buy credits. At OpenAI you need permission to manage billing for that API organization.
2. To keep your tier history, replace the card inside the organization or billing account you already use instead of opening a new account.
3. To give yourself time to fix a decline, set each reload threshold at a few days of peak usage, not a few hours.
4. To keep a limit from looking like an outage, compare each monthly cap on reloads or extra usage with what you really spend in a month. OpenAI, Google, xAI, Cursor and Firecrawl all let you set one.
5. To hear about failures fast, send billing emails to an inbox somebody reads. OpenAI’s billing article describes an email when a recharge fails, and that may be the only warning you get.
6. To size the card limit, add up the reloads you expect in a month across all the vendors on the card. On a card you pay monthly, each top-up adds to the balance until you pay it off.

Here’s how that last step goes wrong. Say your team spends $15,000 a month across OpenAI, Anthropic and Gemini, and each vendor reloads $2,500 at a time. That’s six reloads in a cycle. With a $10,000 limit, the first four go through and the fifth declines, even though none of the charges was big.

Marketplace billing changes who the merchant is. Claude through Claude Platform on AWS is billed through AWS Marketplace, not through Anthropic credit purchases, per Anthropic’s rate limits page. Those charges show up on your AWS bill and earn whatever your card pays at AWS.

Should you use one card or a virtual card per vendor?

One card is simpler. With a virtual card per vendor you can lock or replace one without touching the others, which helps if you have plenty of vendors and someone to look after them.

Ramp offers unlimited virtual cards you can lock to a vendor, each with its own limit per transaction, day or month. Brex suggests a dedicated virtual card for each vendor, so locking or replacing one card only touches that vendor. Slash lists unlimited virtual cards on both its Free and Pro plans, per its pricing page.

Separate cards also keep a decline contained. If one vendor’s card gets frozen, the others carry on reloading. What you pay for that is more cards to watch and more limits to keep up with your spend.

Where does the ernd card fit?

We make the ernd card, so weigh this section with that in mind. It’s a credit card for US companies. Card spend earns 4% back on AI vendor spend, such as OpenAI and Anthropic, and 1.5% back on everything else. Rewards are points, redeemable as cash back.

On the $15,000 a month from the example above, 4% back is $600 a month, or $7,200 a year.

We also have partners across the AI stack, and cardholders get their offers. Wafer, which runs open-weight model inference, gives ernd cardholders 10% back. Hopper gives 8% back on API credits plus a free first month. Agent 37 gives 4% back, Conifer gives 6% back, and Superserve and AgentPhone give 5% back. Firecrawl doubles your usage limits. Trace sells its Startup pen test plan to cardholders for $1,500 instead of $4,000, which is 62% off. Hopper, Agent 37 and Trace pay ernd for their listing. The current list is on ernd.ai/partners.

OpenAI and Anthropic aren’t ernd partners and don’t endorse ernd. The 4% comes from us, not from them.

When is another card the better choice?

When you need more than a good rate on AI charges.

Ramp and Brex are spend platforms. They come with per-vendor virtual cards, spending policies and expense software. The ernd card doesn’t include expense software. Ramp’s AI Spend Management tracks spend at Anthropic, OpenAI, Google Cloud, AWS Bedrock, Cursor and Fireworks AI, and Ramp says you don’t need a Ramp card to use it.

Slash pays up to 1.5% cash back on its Free plan and up to 2% on Pro, which costs $25 a month, with no cap. Purchases in a foreign currency or from non-U.S. merchants earn nothing, and the same goes for Microsoft and Amazon, so an Azure or AWS bill earns no cash back there. Those rules are on Slash’s cashback page.

Amex Business Gold gives up to $300 back in statement credits per calendar year on U.S. purchases of ChatGPT Business subscriptions, with enrollment. It earns 4X points on your top 2 categories each month, on the first $150,000 a calendar year, and it has a $375 annual fee, per the Business Gold page. Software only counts when you buy it directly from U.S. software and cloud system providers, and app purchases through aggregator platforms don’t count, per Amex’s category rules. If ChatGPT Business seats are a big line for you, weigh that $300 credit against the $375 fee and what the card earns on the rest of your spend.

You can also split it. ChatGPT Business goes on Amex for the credit, and the API top-ups go on a card that pays more on AI vendors.

For how prepaid credits, invoices and tax work at OpenAI and Anthropic, read how OpenAI and Anthropic API billing works. There are more card rates in business credit cards for AI spend. If your bill includes open models like DeepSeek or Kimi, open-weight inference pricing covers those providers. Request the card at app.ernd.ai.