Claude Usage Credits Draining Silently: Two Causes and Fixes

A subscription plan feels like a fixed monthly number. But Claude has a second channel money can leave through, separate from the subscription fee, and you can go a long time without noticing that it opened.

That is what happened to me. On August 11 I opened my Claude settings and found a number I did not expect. Usage credits spent: $16.29. I had no memory of buying credits or turning them on. The first thing I did was flip the toggle off.

Usage credits

Usage credits are pay-as-you-go charges that let you keep using Claude after you have exhausted the allowance included in a subscription plan like Pro or Max (Anthropic help center).

My first thought was that my weekly limit was nowhere near spent. But the usage screen showed two limits, not one. Overall usage sat at 59 percent, while the premium model (Fable) limit read 100 percent. That one had bottomed out a while ago, and a usage credits toggle I never knowingly enabled had been picking up the tab ever since. What burned turned out to be promotional credits, so no card was charged. Had auto-reload also been on, everything past that point would have been real money.

The screenshots here are from a Korean-language interface, since that is my own account. The layout and the position of every control are identical in English, and the numbers read the same.

Looking into it, I found this was not just me, and that it happens in two distinct ways. One is what happened to me: a limit genuinely runs out and credits quietly take over. The other is stranger. Claude Code stops mid-task and demands credits while your limits are still perfectly healthy. There are reports from people who bought credits to clear it, watched the purchased credits burn, and still got the same error. This piece starts from the vocabulary, then covers how to tell the two cases apart, what to switch off in each, and where the bug actually came from. Everything below reflects August 2026.

The words: tokens, context window, auto-compact, credits

Four terms carry the rest of this article.

Tokens are the unit Claude counts text in. A short English word is roughly one token.

The context window is how much Claude can hold in mind at once within a single conversation. Think of it as a workbench. The standard bench is 200,000 tokens; the extended one is 1 million.

Compacting clears space on a full bench by summarizing older turns. You can run it yourself with /compact, or Claude Code can fire it on its own as you approach the ceiling. That automatic version is auto-compact, and it sits at the center of the second case below.

Usage credits are a second wallet, separate from your subscription allowance. Leave the toggle on and money comes out of it the moment a limit runs dry.

Two ways credits drain without you noticing

• Case 1. Limit exhausted. A subscription limit genuinely runs out and an enabled credits toggle picks up the charge. Nothing interrupts you, so you only discover it later by reading the spend. This is the one that hit me.
• Case 2. The bug. Your limits are fine, but Claude Code halts and asks you to enable credits anyway. This is not something you did wrong; reports have been accumulating.

Case 1. Limit exhausted: credits picking up the tab

This is exactly what happened to me, and it starts with model-specific limits. Fable, Claude's top-tier premium model, counts against its own weekly limit. So your overall allowance can look healthy while the Fable limit is already gone. Add an enabled credits toggle and credits step in to cover the exhausted limit.

There is one more wrinkle. Anthropic occasionally grants promotional credits to subscribers, and the notification email says they apply automatically with no action needed. On my account the toggle was on as well. So credits can be draining even if you never remember switching anything on.

The fix: three things to check on the usage screen

All three live on the same settings screen, which makes this quick.

1. The usage credits toggle. Turn it off unless you meant to enable it. Switching it off does not affect normal subscription use; you simply wait for the reset when a limit runs out.
2. Auto-reload. This buys more credits with your card when the balance drops. It is the last gate between silent credit drain and an actual charge, so confirm it is off.
3. Both limits. Check the overall limit and the per-model limits separately. If you lean on premium models, that limit fills first.

Case 2. The bug: "Usage credits required"

GitHub keeps collecting the same report. Claude Code stops abruptly and asks for credits while the limit is nowhere near spent. Others bought credits to clear it and found the purchase burned with the error still in place. If you have checked the settings above, your limit is fine, and you are staring at this, you are in case 2.

API Error: Usage credits required for 1M context

Nothing you clicked caused this, and paying will not reliably clear it. The official documentation states that Sonnet 5's 1M window needs no credits, yet the message keeps appearing in practice. Anthropic has not acknowledged it, but if your limits are intact this is most likely a bug. Do not start with the payment button.

The origin of the bug is covered further down; the fixes come first.

Fix 1: check your model, version, and remaining limit

1. Confirm which model you are on. Run /model and read the name on screen. "Sonnet 5" means the default-model bug this article covers. "Fable 5" is a separate premium-model issue, resolved by updating to v2.1.233 or later and signing in again.
2. Check your version. Run claude --version and, if it is old, claude update. Sonnet 5 is only properly supported from v2.1.197 (official documentation). The errors reference says this error is handled automatically when it appears mid-conversation during auto-compact and that no action is required, but it has been re-reported on v2.1.197 and later, so a current version is no guarantee.
3. Confirm you have not actually run out. Run /usage to see what is genuinely left. If your limit is healthy and the error still fires, this is not a limit problem.

Fix 2: disable 1M, tune auto-compact, reselect the model

The official documentation offers three approaches. Note that the page frames them as steps for people who selected a [1m] model deliberately, so they read more as general 1M guidance than as an official answer to auto-compact demanding credits on its own. Still, users blocked mid-compact report getting past it these three ways.

• Turn the 1M window off entirely. Setting CLAUDE_CODE_DISABLE_1M_CONTEXT=1 removes the 1M option from the model picker and pins you to 200,000 tokens. If you set it with export in your shell, it only takes effect after restarting Claude Code (official documentation). Exporting it in a terminal while an already-stuck session runs changes nothing, so relaunch claude afterward. Putting it in the env block of ~/.claude/settings.json applies to running sessions as soon as you save, per the docs. Restarting the session anyway is the safer habit.
• Trigger auto-compact earlier. /autocompact 500k summarizes before you reach that point. Sonnet 5 defaults to roughly 967,000 tokens, so lowering it keeps you away from the 1M boundary where the trouble starts. The trade-off is real: compaction runs more often, older context gets squeezed sooner, and each compaction is itself a call that costs tokens. Match it to how long your sessions actually run rather than setting it as low as possible.
• Reselect with /model sonnet. The official advice is to pick the plain version without the [1m] suffix. In practice plenty of people report this not working. Some got the same error after naming the model explicitly; others found /model had no effect at all inside the VS Code extension. The errors reference itself suggests checking whether a 1M model ID is still sitting in a settings file somewhere. If one attempt fails, move on.

One community workaround comes up often: switch briefly to /model haiku to let auto-compact through, then return to /model sonnet. This is not official guidance, just something multiple users report working in issue threads, so treat it as a stopgap.

Habits help too. Starting fresh with /clear when you move to unrelated work is what the official cost guidance recommends. Save higher tiers like Opus or Fable for work that genuinely needs the judgment. Fable in particular draws on its own weekly limit, which is what leads straight into case 1.

Where case 2 came from: auto-compact since Sonnet 4.6

So how did this start? The 1M window used to be a paid option. Through Sonnet 4.6 it cost credits on every plan including Max (the extended context table). In late June 2026, when Claude Code switched its default model to Sonnet 5 for Pro, Team Standard, and enterprise plans, the 1M window became free by default for the first time (release notes). Auto-compact never caught up with that change. When automatic summarization fires, it reaches for a 1M model regardless of what you selected, applies the old rules, demands credits, and leaves the session dead until you /clear (the original report).

One detail is worth pinning down. This is not a bug Sonnet 5 introduced. The first reports were filed about a month before the switch, on Sonnet 4.6. Back then 1M really was paid, so asking for credits was technically correct; the actual defect was auto-compact jumping to a 1M window nobody requested. That same pattern is now repeating on Sonnet 5, where 1M is free.

A separate case is worth separating out. Fable 5's credit demands came from a different bug, where an expired sign-in token left the client unable to verify subscription tier. Anthropic closed that one on August 17 with an explanation, and the fix is step 1 above. The three core Sonnet reports, by contrast, were still open as of August 20. One was folded into a tracking issue on August 19, but that was consolidation rather than a fix, and no official answer on cause or timeline has followed. Until then, the checks and settings above are what you have.