Claude Code just got a new control that changes how it thinks.
Anthropic recently introduced the Effort parameter — a way to control how many tokens Claude uses when responding. You can now trade off between response thoroughness and token efficiency with a single setting.
If you used Claude Code before this update, you were always running at High effort. Every task, whether a quick file search or a complex architectural refactor, got the same thinking budget. That's wasteful. This update fixes that.
What Is the Effort Parameter?
The Effort parameter controls how eager Claude is to spend tokens when responding to your request. More tokens generally means deeper reasoning, better code structure, and more thorough analysis — but also higher cost and slower response.
It's available on all supported Claude models with no beta header required. You can use it directly in Claude Code or through the API.
One important note: setting Effort to High produces exactly the same behavior as not setting it at all. High is the current default.
The Four Effort Levels
Low — Speed and Cost First
The most token-efficient setting. Claude reduces its reasoning depth significantly to save tokens and respond faster.
Best for: sub-agents, simple tasks, file lookups, quick questions where you need the lowest cost and fastest response.
Medium — The Sweet Spot
A balanced approach with moderate token savings. Claude applies enough reasoning to handle complex agentic tasks without going overboard.
Best for: most everyday Claude Code sessions — agentic workflows, code reviews, feature implementations where speed and quality both matter.
High — The Current Default
Full capability, equivalent to how Claude Code worked before this parameter existed. Complex reasoning, difficult coding problems, extended agentic tasks.
Best for: any task you'd run normally. This is what you've always been using.
Max — Deepest Reasoning, No Limits
Opus 4.6 only. No constraints on token spending. Claude thinks as thoroughly as it possibly can.
If you try to use Max with any other model, it returns an error. You must switch to Opus 4.6 first.
Best for: tasks requiring the deepest possible reasoning — complex refactors, debugging hard-to-reproduce bugs, architectural decisions, anything where quality is worth the token cost.
How to Set It in Claude Code
Type /effort in Claude Code and you'll see all available options: low, medium, high, max, and auto.
If you choose auto, Claude decides which effort level is most appropriate based on the task you give it.
The setting applies to your current session only.
The Live Demo: Same Prompt, Different Results
To show the real difference, I ran the exact same prompt — "create a landing page for my digital product" — twice. Once with Medium effort, once with Max effort.
Medium Effort Result
Claude generated a single index.html file with everything embedded — HTML, CSS, and JavaScript all in one place. The page looked good. Animations worked. Links worked. There was one minor layout issue with accordion widths that could be improved, but overall it was solid and functional.
Max Effort Result (Opus 4.6)
Claude created a proper project structure:
product-landing/
├── index.html
├── css/
│ └── style.css
└── js/
└── main.js
Not just better structure — better output. The layout was tighter, the accordion issue from the Medium version was automatically fixed, fade-in animations were added, and the pricing section had a working toggle between annual and monthly pricing. That feature simply wasn't in the Medium version.
The structural difference alone makes the reasoning visible. Max effort doesn't just think harder — it plans better.
When to Use Each Level
| Task | Recommended Level |
|---|---|
| Sub-agents in a multi-agent system | Low |
| Quick questions, file lookups | Low or Medium |
| Everyday development tasks | Medium |
| Complex feature implementation | High (default) |
| Hard debugging, complex refactors | Max |
| Architectural decisions | Max |
| Cost matters more than perfection | Medium or Low |
The Token Warning
Max effort can consume your daily token quota very quickly. If you're on the Pro plan, be selective. Use Max for tasks where the difference in output quality genuinely matters — complex refactors, architectural work, debugging sessions that have stumped you.
For everything else, Medium will serve you well and save your budget for when you really need it.
Bonus: 1 Million Context Window
Also spotted during this session: Claude Code now shows a 1 million token context window. At the time of recording, this is only available on the Max plan and above. Hopefully it comes to Pro plans in the future.
The Practical Takeaway
- Low = quick draft mode
- Medium = professional work mode
- High = current default (unchanged)
- Max = architect mode — when you need the best possible output regardless of cost
Set your effort level based on the task, not as a one-size-fits-all setting. That's what this parameter enables.
Watch the full demo: Claude Code Just Got Smarter — The Effort Parameter Explained
Top comments (0)