What Is Claude xhigh Effort? April 2026 Guide

xhigh is the fourth reasoning effort level Anthropic added to Claude Opus 4.7 on April 16, 2026. It’s a knob that tells the model to spend more tokens planning before it takes action. For agentic work — Claude Code, MCP tool loops, computer use — xhigh delivers meaningfully fewer wasted tool calls. For quick chat, it’s overkill.

Last verified: April 20, 2026

The short version

Effort	Reasoning tokens	Best for
low	Minimal	Simple chat, quick edits
medium	Moderate	Code review, single-file edits
high	Extended	Multi-file refactors, research
xhigh (new)	Extended + longer planning	Long agentic loops, computer use, complex PRs

Anthropic positions xhigh as “the same reasoning quality as high, but with more planning before tool calls.” In practice it’s also better at when to use tools, not just which to use.

What Anthropic says xhigh actually does

From the Opus 4.7 model card and launch blog:

1. Longer internal deliberation before the first tool call on multi-step tasks
2. More explicit step-by-step decomposition visible in thinking traces
3. Better “I don’t know, let me check first” behavior — less hallucinated confidence
4. Dynamic reallocation — the model can spend more effort mid-task on harder sub-problems

Hexagon (an Anthropic enterprise partner) reported that Opus 4.7 at low effort performs roughly equivalently to Opus 4.6 at medium, and xhigh adds on top of that baseline shift.

When xhigh is worth it

✅ Use xhigh for:

• Long Claude Code sessions — cross-file refactors, adding features that touch 5+ files
• MCP agent loops — anything with 5+ tool calls
• Computer use tasks — OSWorld-Verified-style browser / desktop automation
• Financial analysis — multi-document reasoning, reconciliations
• Bug hunts — “find why this test flakes” kinds of problems
• Security research — Mythos Preview leads here, but xhigh makes Opus 4.7 a close second

⚠️ Probably overkill for:

• One-liner chat messages
• Single-file code edits
• Simple Q&A against your codebase
• Any task that usually finishes in under 30 seconds

For chat-speed interactions, high is still the right default.

Cost: what actually changes

Per-token pricing is identical to high:

Tier	Input $/1M	Output $/1M
Opus 4.7 (any effort)	$15	$75

But xhigh uses more output tokens. In our internal tests on a batch of 20 real Astro tickets:

Effort	Avg output tokens	Avg $/task	Success rate
medium	3,400	$0.26	65%
high	6,800	$0.51	82%
xhigh	11,200	$0.84	91%

So xhigh costs ~1.6× more than high but delivers ~10 percentage points higher success. On agentic tasks, that’s almost always a win because failed tasks cost time, not just tokens.

Claude Code subscribers don’t pay extra. Pro ($20/mo), Max ($100/mo), and Team plans all include xhigh at no additional cost — and Claude Code now defaults to xhigh.

How to turn xhigh on

Anthropic API

from anthropic import Anthropic

client = Anthropic()
response = client.messages.create(
    model="claude-opus-4-7-20260416",
    thinking={
        "type": "enabled",
        "effort": "xhigh"  # low | medium | high | xhigh
    },
    max_tokens=16000,
    messages=[{"role": "user", "content": "Your prompt here"}],
)

Claude Code

xhigh is already the default as of April 16, 2026. To switch:

# Use high for quicker chat-style sessions
claude --effort high

# Explicit xhigh (default anyway)
claude --effort xhigh

You can also set it per-session in /settings effort xhigh.

Cursor, Windsurf, Cline

Each updated to support xhigh between April 16 and April 18, 2026. Look for a “Reasoning effort” dropdown in model settings. If your IDE hasn’t updated, upgrade to the latest version.

xhigh vs GPT-5.4 “xhigh” — are they the same?

No. GPT-5.4 uses reasoning.effort = "high" as its top tier (with a hidden “max” for Pro). Anthropic’s xhigh and OpenAI’s high are not directly comparable — they map to different internal budgets. Cross-provider benchmarks are the only honest comparison.

On the benchmarks we care about:

Benchmark	Opus 4.7 xhigh	GPT-5.4 high
SWE-bench Verified	87.6%	84.1%
SWE-bench Pro	64.3%	57.7%
MCP-Atlas	77.3%	67.2%
Terminal-Bench 2.0	78.0%	75.1%

xhigh earns its name on agentic benchmarks.

Common gotchas

• Streaming latency — xhigh’s longer planning means the first token can take 4–8 seconds. Show a “thinking…” indicator in your UX.
• Max tokens — you’ll want max_tokens ≥ 16000 for xhigh to have room; 8K is often not enough for long agent traces.
• Prompt caching still works — and actually saves more money with xhigh because reasoning tokens aren’t cached but input is.
• Not available in Haiku / Sonnet — xhigh is Opus 4.7 only as of April 2026.

Verdict

xhigh is the new default for agentic work. If you’re using Claude Code, MCP agents, or any multi-step tool loop, switch to xhigh — the extra tokens pay for themselves in fewer failed tasks. For chat-speed interactions, stay on high.

The bigger story: Anthropic is pricing xhigh at the same per-token rate as high, and Claude Code is including it for free. That’s a signal that Anthropic wants xhigh to be the default for agent workloads — and on current benchmarks, they’re right to.