How to buy extra usage on Claude, and the one path that means you usually do not have to

Two pools, not one

Every paid Claude plan has two buckets behind it. The included plan allowance, which Anthropic refreshes on a fixed schedule (weekly on Pro and Max, monthly on Team and Enterprise), and the usage credits pool, which is the pay-as-you-go balance you load yourself. The buy page only manages the second one. Understanding that distinction is the difference between adding funds because you genuinely need to and adding funds because a tool you installed last week is quietly consuming the wrong pool.

The plan allowance is what gets consumed when you chat on the claude.ai web app, the iOS app, the macOS Claude Desktop app, or any other client that authenticates through your OAuth session. No per-token cost is shown because there is not one, you are inside the flat subscription rate. When the allowance runs out, Anthropic tells you to wait for the next reset or, if you have credits enabled, to opt into using them.

Credits, on the other hand, are decremented by anything that authenticates with an Anthropic API key. That includes Claude Code, Cursor, Windsurf, Zed, every ACP wrapper that ships with a bundled key, and any custom code that pulls ANTHROPIC_API_KEY out of the environment. The rate is the published API rate card. There is no subsidy.

What credits are billed at

Usage credits are charged at standard Anthropic API rates. The published 2026 rate card, from platform.claude.com/docs/en/about-claude/pricing:

Opus 4.6 and Opus 4.7 land at $5 per million input tokens and $25 per million output tokens. Opus 4.7 uses a new tokenizer that can represent the same text in up to 35% more tokens, so the same sentence costs slightly more than it did on 4.6. Cache reads are charged at 0.1x the base input rate, the Batch API is 0.5x, and both stack with the rest of the rate card.

A $10 credit therefore covers, roughly, two million Haiku 4.5 input tokens, or three hundred thousand Sonnet 4.6 input tokens, or two million Sonnet output tokens. A heavy Claude Code session that takes ten file-reading subagent calls and a long Opus reasoning pass can chew through a few dollars in a single afternoon. Nothing about the buy flow tells you this, you find out by watching the balance fall.

The buy page is downstream of an architecture choice

Every guide on adding funds to Claude stops at the buy flow. Almost none of them mention that the choice to be on the buy page at all is downstream of how your tooling authenticates. If you spend $200 a month on credits because Claude Code is on every day, the credits are not the problem. The problem is that Claude Code, being a CLI that ships an API key by default, has no path to your plan allowance. It is API-key-only from end to end.

The alternative is a client that authenticates with the same OAuth session you signed in with at claude.ai. From Anthropic's side, that traffic is indistinguishable from a chat in the web app. It decrements the plan allowance first. Credits only get touched once the allowance is exhausted, the same way they only get touched from the web app if you opt to use them.

Fazm is built that way. The relevant Swift is short enough to paste here, so you do not have to take my word for it:

// /Users/matthewdi/fazm/Desktop/Sources/Chat/ACPBridge.swift, line 575
switch mode {
case .personalOAuth:
    env.removeValue(forKey: "ANTHROPIC_API_KEY")  // line 576
case .bundledKey(let apiKey):
    env["ANTHROPIC_API_KEY"] = apiKey             // line 578
}

That is the whole mechanism. When the bridge is in personalOAuth mode, the Node subprocess is launched with no API key in its environment. Without that variable, the bridge falls back to the cached OAuth tokens from your claude.ai sign-in. Every subsequent request is attributed to your user, not an API consumer, and the plan allowance pays for it.

New Fazm users get a small built-in trial allowance first. The cap is declared at ChatProvider.swift:845 as static let builtinCostCapUsd: Double = 10.0. Every Claude response returns a costUsd field that gets added to builtinCumulativeCostUsd, and the moment the running total crosses $10 the app calls await switchBridgeMode(to: "personal") at line 4281 and flips to your OAuth session. From that point on, you are spending your plan allowance, not the credit pool.

When buying credits is still the right answer

OAuth is the right answer for Mac-side automation, chat, and any workflow you can run through a client that supports it. There are three cases where buying credits is the honest answer:

• You depend on Claude Code, Cursor, or Windsurf for daily work and the credit spend is acceptable as a tool cost.
• You are building your own integration against the Anthropic API and need credits to test against the real model.
• You blew through the plan allowance on the web app and need to keep working today without waiting for the next reset.

In all three cases the buy flow at the top of this page is the same. The difference is that in the first case the smarter answer is often to run a wrapper that uses OAuth and only fall back to credits when the plan is exhausted, which is the architecture Fazm ships.

Frequently asked questions