guide

Unlock Claude Code's 1M Token Context Window: Two Lines in settings.json Eliminate Auto-Compaction

AI Tools Hub
#Claude Code #Anthropic #1M context #Claude Sonnet 4.6 #configuration

Set ANTHROPIC_DEFAULT_HAIKU_MODEL and ANTHROPIC_DEFAULT_SONNET_MODEL to claude-sonnet-4-6-1m in .claude/settings.json to run all Claude Code tasks on the 1M token context window. Build an entire SaaS in one session without auto-compaction interrupting your flow.

A two-line addition to .claude/settings.json expands Claude Code’s context window from 200K to 1M tokens, eliminating auto-compaction during long coding sessions.

The Configuration

Add the following to .claude/settings.json (create if it doesn’t exist):

{
  "env": {
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "claude-sonnet-4-6-1m",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-4-6-1m"
  }
}

After adding this, restart Claude Code and switch to the Sonnet model for the settings to take effect.

What Changes

Claude Code internally routes tasks across multiple models. Haiku handles sub-tasks and lightweight processing; Sonnet handles the primary reasoning work. These environment variables override those defaults with claude-sonnet-4-6-1m, which means all tasks run on the 1M token context window.

In standard operation, Claude Code runs auto-compaction as codebases grow—summarizing and truncating older conversation history and code references. With 1M context enabled, this compaction threshold moves dramatically further out, allowing you to maintain the full development context across long sessions.

Practical effects:

  • Reference codebases with thousands of files simultaneously
  • Retain all change history and decision context throughout a session
  • Build entire SaaS projects in a single session without context loss
  • Eliminate the “memory loss” problem from mid-session compaction

Technical Background

According to Anthropic’s official documentation, Claude Opus 4.6 and Sonnet 4.6 support 1M token context windows via the context-1m-2025-08-07 beta header. The claude-sonnet-4-6-1m model ID is a dedicated endpoint that automatically applies this beta header.

1M token context translates to approximately:

  • 750,000 words (English)
  • 3.4 million Unicode characters

Cost Implications

An important caveat: Anthropic applies long context pricing to requests exceeding 200K tokens. Extended sessions will cost more than standard usage.

Cost-effectiveness considerations:

  • Short tasks and simple queries: 200K context is sufficient, no additional cost
  • Large codebase long-form development sessions: 1M context provides clear value
  • Cost-conscious users: Combining with the multi-model optimization strategy covered yesterday—applying 1M context only where it’s genuinely needed—is the rational approach

Setting File Location

.claude/settings.json can be placed in the project root (per-repository settings) or home directory ~/.claude/settings.json (system-wide default).

If the settings file doesn’t exist:

mkdir -p ~/.claude && cat > ~/.claude/settings.json << 'EOF'
{
  "env": {
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "claude-sonnet-4-6-1m",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-4-6-1m"
  }
}
EOF

Summary

This is a practical technique made viable by the 1M context support in Claude Sonnet 4.6. By overriding Claude Code’s internal model assignments, it significantly reduces one of the biggest obstacles to agentic coding—context loss from mid-session compaction.

For developers working on large-scale projects, particularly those who need sustained coherence across long development sessions, this configuration is worth testing. The long context pricing is the primary tradeoff to evaluate against your use case.

Related articles:

Source:

  • Anthropic official model documentation
  • Claude Code settings reference

関連記事

人気記事

Comparison

ChatGPT(OpenAI)とClaude(Anthropic)の機能比較 2026年版。コーディング・長文解析・コスト・API料金の違いを検証

ChatGPT(GPT-4o/o3)とClaude(Sonnet 4.6/Opus 4.5)を2026年時点の最新情報で比較する。コーディング能力、長文処理、日本語品質、API料金、無料プランの違いをSWE-benchなどのベンチマーク結果とともに解説する。

続きを読む →
opinion

【2026年2月20日 所感】「AIがコードを書く」は仮説から現実になった——しかし私たちはその意味をまだ消化できていない

2026年2月20日に観測したコーディングエージェント関連ニュースの総括と所感。Anthropicの自律性研究、cmux、MJ Rathbunのエージェント事故、HN「外骨格 vs チーム」論争、Stripe Minions週1000件PR、Taalas 17k tokens/sec——朝から夜までの流れを通じて見えてきた「AIがコードを書く時代」の実相を考察する。

続きを読む →
tool

868のスキルをnpx 1コマンドで——「Antigravity Awesome Skills」が主要AIコーディングエージェントの共通スキル基盤になりつつある

Claude Code・Gemini CLI・Codex CLI・Cursor・GitHub Copilotなど主要AIコーディングアシスタントを横断する868以上のスキルライブラリ「Antigravity Awesome Skills」(v5.4.0)を詳細分析。Anthropic・Vercel・OpenAI・Supabase・Microsoftの公式スキルを統合した設計思想、ロール別バンドル・ワークフロー機能、SKILL.mdによる相互運用性のアーキテクチャを解説する。

続きを読む →

最新記事

0 tools selected