Claude Code usage limits: how to stretch them
Running out of Claude Code usage early in the day is a common complaint. Anthropic's help article Models, usage, and limits in Claude Code (dated 15 April 2026) explains how usage is metered and what drives it. This guide summarises that article and keeps Anthropic's advice separate from our own notes. We have not run measurements ourselves, and Anthropic says model names and availability change, so treat /model in your own account as the source of truth.
How usage is metered
| You signed in with | What you get, per Anthropic | What running out looks like |
|---|---|---|
| A Claude Enterprise seat | A pool of usage included in your organisation's plan, reset on a rolling window. | A "limit reached, resets at time" message. |
| An API key (Console, Bedrock, Vertex or Microsoft Foundry) | Pay-as-you-go, billed per token to that account. | No hard stop. The account is charged for what it uses. |
The same article says /cost shows your running spend for the session when you use an API key. It does not describe Pro or Max plan limits on this page. Anthropic covers those in a separate help article, Use Claude Code with your Pro or Max plan. Our Claude Code record lists the plans and prices as published.
What uses tokens
Anthropic says every turn sends the conversation so far, your project context (CLAUDE.md and files Claude has read) and your new prompt. The conversation grows fastest, so a long session carries everything it has read and changed on every later message. Anthropic says this is where both cost and context limits come from.
Five habits Anthropic recommends
- Clear between tasks. Run
/clearwhen you switch to an unrelated task. Anthropic calls it the most effective lever for quality and cost. It cannot be undone, and CLAUDE.md and project files stay. - Match the model to the job. Anthropic describes Sonnet as the default for most coding, Opus for harder problems at meaningfully more quota, and Haiku for quick mechanical work. A common pattern is to plan with Opus and execute with Sonnet, which
/model opusplandoes. - Point at files instead of pasting them. Pasted text stays in context for the session. Anthropic notes that the @ prefix injects the whole file plus its CLAUDE.md tree, so use a bare path when saving tokens, and trim logs to the relevant 20 or 30 lines.
- Keep CLAUDE.md lean. It is added to every turn. Anthropic suggests adding a note only the second time you correct the same thing, keeping the file under roughly 200 lines, and pruning stale notes every few weeks.
- Ask for a plan before big changes. For anything touching more than two or three files, use Plan Mode or ask for the list of files and edits first, correct it, then execute.
Context window versus usage limit
- A full context window is different from a usage limit. Anthropic says to use
/compactto summarise and keep going, or/clearif the old history is not needed. It also says Claude Code auto-compacts near the limit. - If you hit a limit on an Enterprise seat, the message gives the reset time. Anthropic says you can switch to a lighter model, or fall back to an API key if your organisation allows it.
- With an API key, Anthropic says unexpectedly high spend almost always traces back to very long sessions that were never cleared.
Our notes
These are our reading of the article, not Anthropic's claims. The advice is about keeping what is resent each turn small, which is why clearing, short context files and plans help. We have not measured how much each habit saves. If you are choosing between coding agents on cost, see Replit Agent vs Claude Code vs Devin and Devin pricing explained.
Source: Anthropic Help Center, Models, usage, and limits in Claude Code. Spot a change? Tell us.