Unlock Claude Code's 1M Token Context Window: Two Lines in settings.json Eliminate Auto-Compaction
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:
- Claude Sonnet 4.6 Release: Opus-Level Intelligence at Sonnet Price
- Cut OpenClaw API Costs by Up to 90%
Source:
- Anthropic official model documentation
- Claude Code settings reference
Related Articles
Strict Linting Dramatically Improves LLM Code Quality: TypeScript Go, Oxlint, and Oxfmt for the AI Era
Meta engineer and OpenClaw developer Christoph Nakazawa's technical article demonstrates that strict guardrails—TypeScript Go's 10x faster type checking, Oxlint, and Oxfmt—significantly improve LLM code output quality, with GPT 5.2 Codex experiments showing fewer bugs under strict configurations.
February 2026 AI Tools Guide: Best Picks for Development, Business, and Creative Work
Answer to 'Which tool should I actually use?' 15 AI tools selected based on popularity and proven results, categorized for developers, business users, and creators. Includes Claude Sonnet 4.6, Cursor, OpenClaw, and more standout tools of February 2026.
Cut OpenClaw API Costs by Up to 90%: A Practical Multi-Model Strategy
Running OpenClaw on a single frontier model burns money on routine tasks. A multi-model routing strategy, prompt caching, and local models can reduce API costs by 80-90% while maintaining output quality for complex tasks.
Popular Articles
Claude Code v2.1.93 Released - Deferred Permission Decisions, Flicker-Free Rendering, and More
Anthropic releases Claude Code v2.1.93 with deferred permission decisions for PreToolUse hooks, flicker-free rendering option, PermissionDenied hook, and named subagent typeahead support.
Claude Code v2.1.92 Released - forceRemoteSettingsRefresh, Bedrock Setup Wizard, and More
Anthropic releases Claude Code v2.1.92 with forceRemoteSettingsRefresh policy setting, AWS Bedrock setup wizard, /cost command improvements, and numerous bug fixes.
Claude Code v2.1.84 Release - PowerShell Tool Preview and Environment Configuration Enhancements
Claude Code v2.1.84 introduces PowerShell tool for Windows, new environment variable overrides for model selection, idle session handling improvements, and various stability fixes.
Latest Articles
Claude Code v2.1.93 Released - Deferred Permission Decisions, Flicker-Free Rendering, and More
Anthropic releases Claude Code v2.1.93 with deferred permission decisions for PreToolUse hooks, flicker-free rendering option, PermissionDenied hook, and named subagent typeahead support.
Claude Code v2.1.92 Released - forceRemoteSettingsRefresh, Bedrock Setup Wizard, and More
Anthropic releases Claude Code v2.1.92 with forceRemoteSettingsRefresh policy setting, AWS Bedrock setup wizard, /cost command improvements, and numerous bug fixes.
Claude Code v2.1.91 Released - MCP Tool Result Persistence and Improved Edit Tool
Claude Code v2.1.91 introduces MCP tool result persistence override, improved shell execution controls, and enhanced Edit tool efficiency.