Claude Code, Opus 4.7 Token Consumption Surge Issue

After Anthropic recently released Opus 4.7 and automatically switched active Claude Code sessions from Opus 4.6 to 4.7, a surge in token consumption has been reported during this process.

Causes of the token consumption surge:

• Reduced tokenizer efficiency: 4.7's tokenizer uses approximately 1.35x more tokens for the same input.
• Lack of context compression: The previous 4.6 performed auto-compact at a 200K context, but 4.7 (1M version) does not perform this, adding to the cost as conversations get longer.
• Result: Overall token consumption rate increased by approximately 4x, causing the existing 5-hour quota to be exhausted in just 30 minutes.

Solution: Adding the following to the ~/.claude/settings.json configuration file to force-specify the model as Opus 4.6 can resolve the issue.

{
  "model": "claude-opus-4-6"
}

Additionally, according to benchmark results, Context Recall performance also dropped significantly in 4.7 compared to 4.6 (91.9% → 59.2% at 256K).