The 1M Context Trap for Opus - More Context Makes the Model Lazier

The 1M token context window sounds like a dream. Fit your entire codebase in context. No more worrying about what to include. Just dump everything and let the model figure it out.

In practice, it is a trap. The more context you provide, the lazier the model gets.

The Laziness Problem

When you give Opus a 10,000-token context with a focused task description, it produces precise, targeted output. It reads every line carefully because there are not that many lines to read. It understands your intent because the signal-to-noise ratio is high.

When you give Opus a 500,000-token context with the same task, something changes. The model starts skimming. It latches onto the first plausible approach it finds rather than considering alternatives. It misses constraints that were clearly stated but buried in the middle of the context. It produces output that is correct-ish rather than correct.

This is not a bug in Opus specifically. It is a fundamental property of attention mechanisms. As the context grows, the model's attention gets distributed more thinly across more tokens. Important details in a sea of context get less attention weight than the same details in a focused context.

The Practical Impact

Developers who discover the 1M context window often go through a predictable arc:

1. Excitement - "I can fit everything in context!"
2. Experimentation - dump the whole codebase and ask for changes
3. Disappointment - the output is vague, misses edge cases, and ignores stated constraints
4. Retreat - go back to curating context carefully

The sweet spot is usually between 20,000 and 100,000 tokens of well-curated context. Enough for the model to understand the full picture, but not so much that important details get diluted.

Better Strategies Than More Context

Instead of maximizing context, maximize relevance:

• Scope each agent to specific files - do not include the entire codebase when the task only touches three files
• Use CLAUDE.md for constraints - put the most important rules at the top where they get the most attention
• Split large tasks into smaller ones - each with its own focused context
• Let the agent pull context on demand - tools like grep and file reading let the agent find what it needs rather than you guessing what to include

The counterintuitive truth is that the best context strategy is not "more" - it is "less but better." A 30,000-token context with exactly the right files will outperform a 300,000-token context with everything.

Fazm is an open source macOS AI agent. Open source on GitHub.

More on This Topic

• 200K Context vs 1M and Aggressive Context Clearing
• One Million Token Context vs Scoped Agents
• Stop Fighting the Context Limit - Scope Each Agent