Claude Code xhigh Effort Level Guide — Go Deeper on Agentic Work with Opus 4.7

Every Claude Code power user has been there: "high effort gives me decent results but misses the nuance. max is thorough but slow and expensive. What I need is something exactly in between."

In May 2026, Anthropic filled that gap. xhigh — a new effort level sitting between high and max, exclusive to Claude Opus 4.7.

But xhigh is not the only update. Windows users can now run Claude Code without installing Git Bash. The claude ultrareview command can run non-interactively in CI pipelines. And MCP servers auto-retry on transient connection errors.

As an EdTech CEO who uses Claude Code daily for agentic tasks, here is what each of these improvements means in practice.

Table of Contents

1. What Is the xhigh Effort Level
2. What Changed in Claude Opus 4.7
3. Windows Support — No Git Bash Required
4. ultrareview for CI Automation
5. MCP Auto-Retry and Stability
6. Real-World Scenarios

What Is the xhigh Effort Level

xhigh is an Opus 4.7-exclusive effort level that provides high-intensity reasoning while being faster than max.

Claude Code's effort levels control how deeply Claude reasons through a task. The existing scale was low → medium → high → max. With Opus 4.7, there is now a fifth option: low → medium → high → xhigh → max.

A helpful analogy: imagine a pitcher who throws at full power (max) vs. 80% (high). xhigh is the 90% throw — more effort than high, faster than max.

How to Set xhigh

xhigh can be activated three ways:

# CLI flag
claude --effort xhigh

# In-session command
/effort xhigh

# Model picker (UI)

Note: xhigh is only available on Opus 4.7 with a Max plan subscription. Other models automatically fall back to high when xhigh is selected.

What Changed in Claude Opus 4.7

To feel what xhigh actually unlocks, it helps to understand what is new in Opus 4.7 itself.

High-Resolution Image Support

Opus 4.7 is Anthropic's first Claude model with high-resolution image support. Maximum image resolution increased from 1568px (1.15MP) to 2576px (3.75MP) — about 3.3× sharper.

Why does this matter in practice? When you send Claude a UI screenshot and ask "find the layout issues," the previous resolution limit made fine details blurry and unreadable. Now analysis at near-screen clarity is possible.

1M Context + 128k Output

Opus 4.7 supports a 1 million token context window and 128k max output tokens. For large-codebase reviews, comprehensive document drafts, or multi-file agentic tasks, this means significantly fewer interruptions and breaks in continuity.

More Literal Instruction Following

Opus 4.7 follows instructions more literally than previous models. "Edit only this function" means it will not touch other files. This is a precision improvement — and a reminder that the more explicit your prompt, the better the result.

The more agentic the task, the more precise the prompt needs to be.

Windows Support — No Git Bash Required

Claude Code no longer requires Git for Windows to be installed.

Previously, running Claude Code on Windows required Git Bash. Now, when Git Bash is absent, Claude Code automatically uses PowerShell as the shell tool. This removes a meaningful friction point for Windows users, particularly in enterprise or educational environments where installing Git may be restricted.

ultrareview for CI Automation

claude ultrareview can now run non-interactively from CI/CD pipelines.

Previously /ultrareview only worked inside an interactive session. The new subcommand runs it headlessly from the command line:

# Run ultrareview on a target
claude ultrareview [target]

# Output results as JSON
claude ultrareview [target] --json

Exit codes: 0 on completion, 1 on failure. This makes it straightforward to trigger automated code review when a pull request is opened and block a build if quality thresholds are not met.

# GitHub Actions example
- name: Claude Ultrareview
  run: claude ultrareview ./src --json > review-report.json

MCP Auto-Retry and Stability

MCP servers now automatically retry up to 3 times on transient startup errors.

Anyone running MCP (Model Context Protocol) servers has experienced the frustration: a transient error during initialization disconnects the server, and re-connecting requires manual intervention. Claude Code now retries automatically, which significantly reduces disruptions caused by momentary network issues or slow initialization.

For complex multi-MCP agentic workflows, this improvement in connection stability has an outsized impact on reliability.

Real-World Scenarios

Here are workflows you can apply today.

Scenario 1: Long-Running Agentic Code Task

# Set xhigh for complex refactoring
claude --model claude-opus-4-7 --effort xhigh

# Prompt the agent
> Analyze all authentication logic in this repository,
> identify security vulnerabilities, and for each one
> provide a corrected code snippet with an explanation.

xhigh delivers near-max analysis depth with meaningfully faster response times — the right balance for long agentic tasks where you need thoroughness without the full max latency.

Scenario 2: Educational Content Quality Pipeline

# Auto-review AI-generated lesson materials before publishing
claude ultrareview ./content/lessons --json | python check_quality.py

Before distributing AI-generated educational content, run it through an automated quality gate to catch errors and flag inconsistencies.

Scenario 3: Multi-MCP Agent Workflow

{
  "servers": {
    "notion": { "command": "npx notion-mcp" },
    "filesystem": { "command": "npx fs-mcp" },
    "web": { "command": "npx web-mcp" }
  }
}

With auto-retry, multi-server agentic workflows become more reliable and require less babysitting.

Closing Thoughts

The through-line in this Claude Code update is practical improvement for agentic work. xhigh expands the range of reasoning choices. ultrareview automation enables quality-gated pipelines. MCP auto-retry reduces the fragility of complex workflows.

For anyone practicing vibe coding — rapidly building and iterating with AI — the xhigh + ultrareview combination is worth making your new default setup.

Anthropic described Opus 4.7 as "the most capable model that can be safely deployed broadly." The tooling around it is catching up fast.

Related Posts

• Gemini Has Left the Chat — File Generation, Scheduled Actions & Personalization
• Notion AI Credits Transition — What Changed in May

What is the hardest part of using Claude Code in your workflow? Share in the comments!

Sources:

• What's New in Claude Code — Claude Code Docs
• Introducing Claude Opus 4.7 — Anthropic
• What's New in Claude Opus 4.7 — Claude API Docs
• Claude Code CHANGELOG — GitHub
• Anthropic Release Notes May 2026 — Releasebot