Stretch your Claude Code usage limit (the real 2026 settings)

Why you hit the limit so fast

Claude Code bills by tokens, and tokens scale with context: the more the model has to read, the more each message costs. Two things quietly inflate that context. Your chat history grows every turn, and Opus (the heaviest model) burns tokens far faster than Sonnet. So a long session on the wrong model is the fastest way to a "you've hit your limit" message. The fixes below attack both.

Source: Anthropic, Manage costs effectively.

Setting 1: run the right model with /model opusplan

In a session, type /model and pick opusplan. Anthropic's docs describe it as a special mode that "uses opus during plan mode, then switches to sonnet for execution." So Opus does the thinking (architecture, edge cases, the plan) and Sonnet does the typing (the actual line-by-line edits). You only pay Opus rates for the part that needs the deep reasoning.

/model opusplan

• 🧩 What it is: a hybrid model alias. As of July 2026, opus resolves to Opus 4.8 and sonnet resolves to Sonnet 5 on the Anthropic API.
• ✅ When to use it: any workflow where you use plan mode regularly (refactors, features, anything you'd sketch before building).
• ⚠️ Heads-up: it only shines if you actually plan first. If you skip planning and just fire one-off edits, you never get the Opus phase and it behaves like plain Sonnet. To force the plan step, press Shift+Tab to enter plan mode before you build.

Source: Anthropic, Model configuration.

Setting 2: keep context small with /compact and /clear

The single biggest token drain is a bloated conversation you never reset. Claude Code helps two ways.

Auto-compaction already runs on its own: when your session gets close to the model's context limit, Claude summarizes the older history to free up space. That "context left" or "compacting" warning is not a usage limit, it is this kicking in.

You can also drive it manually, and this is where people leave value on the table:

/compact Focus on the code changes and API usage
/clear

• 🗜️ /compact squeezes the current thread into a tight summary so you keep the important context and drop the noise. Add a focus instruction (like above) and Claude keeps exactly what you tell it to. You can even set a default in CLAUDE.md so every compact preserves the right things.
• 🧼 /clear wipes context entirely when you switch to unrelated work. This is the one most people forget. Stale context from your last task is dead weight on every new message. Run /rename first if you want to /resume that session later.

Source: Anthropic, Reduce token usage.

Setting 3: control thinking depth with /effort (and ultrathink)

This is the tip that changed the most, so here is the current truth. Extended thinking is now on by default in Claude Code because it genuinely improves hard tasks, and thinking tokens are billed as output tokens. That means deep reasoning is no longer something you switch on, it is something you dial up or down.

The real lever now is /effort. It sets how hard the model reasons for the whole session, from low to max.

/effort low
/effort high
/model   (then use the left/right arrows to slide effort)

• 📉 To save tokens: drop to /effort low or medium for short, simple, not-very-brainy tasks. On Opus 4.8 and Sonnet 5 the default is high, with low, medium, high, xhigh and max available.
• 📈 For one hard task only: add the word ultrathink anywhere in that prompt. Per Anthropic's docs it "requests deeper reasoning on that turn without changing your session effort setting." So it is a one-turn boost, not a permanent mode.
• ⚠️ Correction worth knowing: the old magic phrases like "think hard", "think harder" and "megathink" are no longer special keywords, they are now read as ordinary text. Only ultrathink is still recognized. If you learned the tiered think-words trick from an older video, that part is stale.

Source: Anthropic, Adjust effort level.

3 bonus moves to stretch it even further

• 🧰 Trim MCP overhead. Every connected MCP server can add tool definitions to your context. Run /context to see what is eating space and /mcp to disable servers you are not using right now. Prefer CLI tools (gh, aws, gcloud) where you can, they cost no per-tool listing. Source.
• 📎 Move heavy CLAUDE.md rules into skills. Your CLAUDE.md loads into context at the start of every session, so long instruction blocks are paid for even on unrelated work. Anthropic suggests keeping it under about 200 lines and moving specialized workflows into on-demand skills. Source.
• 🔍 Delegate the noisy stuff to subagents. Running tests, fetching docs, or grepping huge logs can flood your main context. Hand those to a subagent so only the summary comes back, and set simpler subagents to model: haiku. Source.

How to get the most out of it

When NOT to reach for these