How to Add Credits to the Anthropic / Claude API When Your Card Is Declined (2026)
Hitting 'Add credits' in the Anthropic Console and watching your card get declined? Here's why it fails, how Claude API billing actually works (prepaid credits), and the two fixes that work.
1. The API is not the same as a subscription
A claude.ai subscription (Pro/Max) covers usage in the web app and in Claude Code. The API at console.anthropic.com is a separate, per-token developer interface with its own balance — building an app, running batch jobs, or powering an agent all go through the API, and none of that draws from a Pro or Max subscription. If you're hitting Add credits and getting declined, this is the surface you're on.
2. Why Add credits keeps declining
Console billing runs through a standard card-payment processor, and it screens each attempt before an authorization is even created. The strongest signal in that screen is the issuing country of the card: cards issued outside the US and EU are declined at a materially higher rate, independent of the balance on the card or how carefully the details were entered. That's why the number goes in cleanly, the form submits, and the charge simply never lands — nothing about the input was wrong. Two consequences follow. Retrying the same card repeatedly doesn't help and can make the account look riskier to the same check; and the fix has to change the one variable the check actually looks at — swap in a card whose issuing country the processor accepts, rather than re-entering the same details a fourth time.
3. What prepaid credits actually mean for you
Anthropic's billing model matters as much as the payment method: it's prepaid credits, not pay-later. You load a balance up front, and every call is metered against it in real time. When the balance reaches zero, the next call doesn't slow down or queue — it fails immediately with an error. Two practical implications follow. In production, put a balance check on a schedule rather than discovering the drop from a support ticket; a job running overnight on a thin balance can die halfway through with no warning. And if you only want to confirm your code path works before committing money, do that on a small top-up rather than loading a large balance for a test that needs a couple of dollars of actual usage.
4. Fix one: add credits with a card the Console accepts
If you need to stay on your own official Anthropic account — for an invoice in your organization's name, an SLA, or because a client requires it — the direct fix is a card whose issuing country the Console's payment check accepts. A US-issued virtual Visa clears that specific screen. Open Console → Settings → Billing, add the card, and use Add credits with a small amount first (around $5) to confirm it clears before loading a larger balance. cocodot issues this kind of card for a flat $9.9, funded from a balance you top up locally — check the pricing page for the current top-up fee, since it's reviewed periodically. One thing this route does not solve: the network path to Anthropic's endpoints is a separate problem from billing, and production traffic run from a country with unreliable direct access commonly still needs its own connectivity plan regardless of how payment gets resolved.
5. Fix two: call Claude through a relay you can fund directly
If the goal is simply to get Claude working in your code — the account doesn't need to be in your own name, and you don't need an official invoice — this route is faster: sign up for an OpenAI-compatible relay that takes a payment method you already have, top up, create a key, and point base_url at it. With cocodot specifically: register, verify your email for a small trial credit, fund the balance, and set base_url to https://cocodot.co/api/ai/v1 with the Claude model code — the current lineup is public at cocodot.co/api/ai/v1/models, so pull it live rather than copying a model name out of any article, since the lineup changes. Nothing else in your integration changes; you're still calling the same underlying Claude models, billed per token, at prices that don't exceed the official list.
6. Claude Code needs Anthropic's format specifically
One detail trips people up when they move to a relay: Claude Code speaks Anthropic's own API shape, not the generic OpenAI format most relays expose. If your integration is a script calling chat completions, an OpenAI-compatible endpoint is all you need and the relay route above works as described. If you specifically want Claude Code itself to point at a relay, confirm the relay publishes an Anthropic-compatible endpoint before you commit to it — cocodot has one, currently in beta, with configuration on cocodot.co/claude-code. Tools that only speak OpenAI's format, like most Cursor or Cline integrations, aren't affected by this distinction and work fine with either kind of endpoint.
7. Picking a model and keeping the bill predictable
Whichever billing path you land on, the same habits keep spend under control. Route everyday tasks to Sonnet or Haiku and reserve Opus for problems that actually need it — capacity and price both scale with model tier, so this is the single biggest lever you have. Turn on prompt caching if your workload repeats context, such as a system prompt, a long document, or a tool schema; it materially cuts the cost of the repeated portion on every call after the first. Set a spend limit in the Console so a bug in a retry loop can't run past a number you chose, and treat your card or relay balance as the backstop underneath that limit, not the primary control.
8. When something still doesn't work
If a card is declined a second time on the same details, check three things before trying a third: the card hasn't expired, the billing address matches exactly what's on file for that card, and you aren't retrying within minutes of the last failure — space it out, since repeated rapid attempts read as riskier rather than more persistent. If credits were charged but don't show as available, check the invoice status on the Billing page before assuming the payment failed; sometimes it's a display lag rather than a failed charge. And if you need a receipt for expense reporting, invoices are downloadable directly from the Console regardless of which card funded them. One more thing worth checking before you conclude billing is broken: confirm the error is actually a payment decline and not a separate connectivity issue presenting the same way, since a timed-out request to Console can look identical to a rejected charge in some client libraries — retry once on a stable connection before treating it as a card problem.
Can't add Claude API credits: which fix is for you
| ① Add credits via official Console | ② Call Claude through a relay | |
|---|---|---|
| Best for | You need the official account, invoices or SLA | You just need Claude working in your code |
| How you pay | Issue a US Visa virtual card → Console → Add credits | Fund the balance directly — no card |
| Account | Your own official Anthropic account | Relay account — still the official models underneath |
| Network | Direct access from China is unstable; still your problem | Direct route, no VPN |
| Setup cost | Issue a card, then the official card-on-file flow | Change one base_url (OpenAI-compatible) |