Hitting Claude usage limits: what to do

A familiar situation: the work is flowing, the agent has hit its stride, and mid-week a message arrives saying your limit is spent. From there you wait for the quota to reset, and nothing you do changes that — money does not solve it, and there is no way to pay for more inside the plan.

Why the limit exists at all

A subscription is a fixed fee for access, not for volume. For that model to work economically, consumption has to be bounded: otherwise a handful of the most active users would absorb everything. Hence quotas and their reset windows.

For irregular work that is convenient: you pay one sum and never count tokens. The problem starts when the load becomes uneven — a large refactor, reading through an unfamiliar codebase, a long agent session. Those are exactly the days that hit the ceiling.

What is usually suggested

Move to a higher plan. Works until you hit the wall again: the ceiling moves, it does not disappear. And you end up paying for your peak rather than your average.

Open a second account. This directly violates the terms of use and ends in a ban. Not recommended.

Wait. Sometimes reasonable — but not when the deadline is today.

Per-token billing instead of a subscription

The alternative is working with an API key against a balance. Then there is exactly one constraint: how much money is in the account. Run out and you top up and continue a minute later, not several days later.

You will have to keep an eye on it, of course: every request costs money, and intensive work drains a balance noticeably. In exchange you pay for precisely what you used, and quiet weeks do not cost you an unused quota.

A practical anchor: one dense Claude Code session runs to hundreds of thousands of input tokens. At $0.30 per million that session costs tens of cents. At Anthropic's list rate for Claude Opus 5.5 of $4 per million input, it is a few dollars. One caveat: vendors charge ~10% of the input price for prompt-cache reads, while Zukko bills cached tokens at the regular price, so for agents with a large repeated context the real gap is smaller than a straight comparison suggests.

How to switch

You set a base URL and a key in the config of whichever tool you use. For Claude Code that is ~/.claude/settings.json:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.zukko.pro",
    "ANTHROPIC_AUTH_TOKEN": "fetch_your_key"
  }
}

Cursor, Codex, Cline and Roo Code are configured the same way — one address changes. Your code and SDK stay untouched, because the API is compatible with the original.

One detail about addresses: Anthropic-compatible clients take no /v1 suffix, OpenAI-compatible ones require it. This is the most common cause of a 404 on first connection.

When a subscription is still better

If your work is steady and fits the quota, stay on the subscription — it is predictable and needs no balance watching. Switching starts to make sense when you hit the limit regularly, or when the load swings: a quiet week, then a marathon.

Many people keep both: a subscription for daily work and a key for the days when the quota is gone and the deadline is today. Switching takes seconds, because it is one line of config.

You can test the second scenario with a minimal outlay: top-ups start at $1, there is no starting balance, and subscribing to the @zukkopro↗ channel credits $0.50 once (link Telegram in the dashboard). That is enough to run a real task and see your own consumption.