What Is Claude xhigh | How to Use the 5-Level Effort System

What Is Claude xhigh — A Reasoning Intensity Setting Added Between high and max

Claude's effort parameter lets you incrementally control how many tokens the model spends generating a response. Increasing the effort level improves the depth and accuracy of reasoning, but also increases processing time and cost.

xhigh was introduced at the same time as the release of Claude Code v2.1.111. It initially supported only Claude Opus 4.7, and was later extended to Claude Opus 4.8. Even after its addition, the API default remains high, so xhigh is not applied unless explicitly specified.

Anthropic's official documentation describes it as follows:

Start with xhigh for coding and agentic use cases, and use high as the minimum for most intelligence-sensitive workloads.

— Anthropic Official Effort Documentation

xhigh targets "long-running agentic coding tasks (over 30 minutes)" and is particularly effective for exploratory processes such as repetitive tool calls, detailed web searches, and knowledge base retrieval.

5-Level Effort Quick Reference

Here is a comparison of the 5 levels defined in the official documentation.

According to the Anthropic Official Effort Documentation, the models that support the effort parameter are: Claude Fable 5 / Claude Mythos 5 / Claude Opus 4.8 / Claude Mythos Preview / Claude Opus 4.7 / Claude Opus 4.6 / Claude Sonnet 4.6 / Claude Opus 4.5 — 8 models in total. "All supported models" in the table refers to this list; Claude Haiku 4.5 is not included (the effort parameter itself is not supported for that model).

Of these, xhigh can only be specified for Claude Fable 5 / Claude Mythos 5 / Claude Opus 4.8 / Claude Opus 4.7 — 4 models. Specifying xhigh for an unsupported model such as Sonnet 4.6 will result in an error. Note that while the official documentation does not explicitly state what happens when effort is specified for a non-supporting model, specifying an unsupported level for a supporting model does result in an error as shown above. Do not assume "it will be silently ignored and still work" — only use model-and-level combinations listed as compatible.

If the parameter is omitted, behavior is equivalent to high. Explicitly setting effort: "high" produces the same result.

Recommended Use Cases for xhigh

The following are the scenarios where Anthropic's official documentation recommends xhigh:

• High-difficulty coding — Implementations requiring deep exploration, such as refactoring across multiple files, fixing an entire test suite, or performance tuning
• Long-running agentic tasks — Complex workflows involving repeated sub-agent calls (tasks taking 30 minutes or more)
• Repetitive tool calls — Tasks where web searches or codebase searches are repeated iteratively to build up an answer step by step
• Detailed knowledge base retrieval — Research-style processing that digs accurate information out of large volumes of documentation

On the other hand, for simple question-answering, text summarization, or lightweight classification tasks that complete quickly, high or medium offers better cost efficiency.

A key principle from the official guidelines: "Use high as the minimum for most intelligence-sensitive use cases, and start with xhigh for coding and agentic use cases."

How to Configure xhigh via the API

To specify xhigh via the API, pass effort: "xhigh" inside the output_config object.

Python example:

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=64000,
    messages=[
        {
            "role": "user",
            "content": "このリポジトリ全体をリファクタリングしてテストを修正してください。",
        }
    ],
    output_config={"effort": "xhigh"},
)

print(response.content[0].text)

TypeScript example:

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic();

const response = await client.messages.create({
  model: "claude-opus-4-8",
  max_tokens: 64000,
  messages: [
    {
      role: "user",
      content: "このリポジトリ全体をリファクタリングしてテストを修正してください。",
    }
  ],
  output_config: {
    effort: "xhigh"
  }
});

Source: Anthropic Effort Documentation

output_config can be specified alongside the thinking parameter. Combining it with Adaptive Thinking is covered in a later section.

How to Configure xhigh in Claude Code

There are two ways to configure xhigh in Claude Code.

Switching during a session with a command:

/effort xhigh

Specifying a flag when launching Claude Code:

claude --effort xhigh

Starting with Claude Code v2.1.154, new sessions on the Max plan default to Opus 4.8 as the model and have /effort xhigh enabled by default (source). Pro plan users and those on older versions still need to specify it explicitly.

The /effort menu in Claude Code also shows ultracode as an option, but this is different from the API effort levels. ultracode is a combined setting that enables xhigh effort plus automatic activation of multi-agent workflows — the value ultracode does not exist as an effort parameter in the API.

Token Consumption and max_tokens Settings When Using xhigh

xhigh consumes significantly more tokens than high. Anthropic's official documentation recommends setting max_tokens to a large value when using xhigh or max, with 64k tokens as a starting point for adjustment.

Key points to keep in mind regarding token consumption:

1. xhigh may consume several times more tokens than high
2. Tasks involving sub-agents or multiple tool calls will consume even more tokens
3. If max_tokens is too small, the model will be cut off mid-reasoning or mid-action, degrading quality

As a practical approach to cost management, measure actual token consumption on your workload before deploying to production, and use staged evaluation — stepping down the effort level based on task complexity — to find the right balance.

How to Combine xhigh with Adaptive Thinking

For Claude Opus 4.7 / 4.8, manual specification of thinking budget via budget_tokens has been deprecated. The recommended approach is now combining thinking: {type: "adaptive"} with effort.

response = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=64000,
    messages=[...],
    thinking={"type": "adaptive"},
    output_config={"effort": "xhigh"},
)

When Adaptive Thinking is enabled at the high, xhigh, or max effort levels, the model will almost always think deeply. At medium or low, it may skip thinking for simple problems.

Note that even if thinking: {type: "adaptive"} is omitted, the effort parameter is still effective. In that case, the effort level influences all tokens other than thinking (text responses and tool calls).

xhigh vs. max — Decision Criteria for Choosing Between Them

The difference between xhigh and max is the balance between token consumption and response quality. Anthropic's official guidelines offer the following approach:

• Use xhigh as the first choice — For most coding and agentic tasks, xhigh provides sufficient quality
• Reserve max for when evaluation metrics show a gap — Only move to max after confirming that xhigh produces insufficient quality
• Avoid overusing max — For structured outputs or relatively simple tasks, max can actually cause excessive reasoning, which may degrade quality

As a practical operational guideline, the cost-effective approach is to establish a baseline with xhigh first, then try max only for tasks where quality is insufficient.

Also note that Claude Opus 4.7 adheres more strictly to effort levels than Opus 4.6. At low or medium, it tends to stay within the instructed scope and avoid unnecessary processing. If reasoning feels too shallow when using low or medium on complex problems, increasing the effort level is more effective than trying to compensate with prompting.

Source: Anthropic Effort Documentation