Claude Code Opus Used Instead of Sonnet (2026)

The Problem

Claude Code ran on Opus 4.6 ($15/$75 per MTok) instead of Sonnet 4.6 ($3/$15 per MTok), resulting in a 5x cost multiplier. A typical 60K-token session that costs $0.36 on Sonnet cost $1.80 on Opus. A heavy session of 200K tokens: $1.20 on Sonnet versus $6.00 on Opus. If the mistake persisted for a full day of 5-10 sessions, the extra cost ranges from $7-$50 in a single day.

Quick Fix (2 Minutes)

  1. Switch model immediately: 
    claude --model sonnet -p "your prompt"
  2. Create a shell alias to prevent recurrence: 
    # Add to ~/.zshrc or ~/.bashrc
alias claude='claude --model sonnet'
  3. Verify current model before expensive tasks: 
    # Check which model is active at session start
# The model name appears in the Claude Code startup header

Why This Happens

Three common causes:

1. Model switched for a previous task and not reverted. A developer uses Opus for a complex architecture task, then forgets to switch back for routine work. Every subsequent session runs at 5x cost.

2. Environment variable override. The ANTHROPIC_MODEL or similar environment variable is set to an Opus model ID, overriding the default for all sessions.

3. Configuration file setting. A global or project settings file specifies Opus as the default model.

The cost impact is immediate and significant:

Session Type Sonnet 4.6 Cost Opus 4.6 Cost Overspend
Quick fix (20K tokens) $0.12 $0.60 $0.48
Bug fix (60K tokens) $0.36 $1.80 $1.44
Feature implementation (120K tokens) $0.72 $3.60 $2.88
Large refactor (200K tokens) $1.20 $6.00 $4.80
Full day (5 sessions, 500K total) $3.00 $15.00 $12.00

The Full Fix

Step 1: Diagnose

Determine how long Opus has been active and the total overspend:

# Check recent session costs
ccusage --sort cost --limit 10
# Look for sessions with unusually high costs
# A 60K-token session costing $1.80+ is a strong Opus indicator
# (Same session on Sonnet would cost ~$0.36)

Check for model configuration overrides:

# Check environment variables
env | grep -i model
env | grep -i anthropic
env | grep -i claude
# Check global settings
cat ~/.claude/settings.json 2>/dev/null | grep -i model
# Check project settings
cat .claude/settings.json 2>/dev/null | grep -i model

Step 2: Fix

# Remove any model overrides from environment
unset ANTHROPIC_MODEL
# Set explicit model in shell profile
echo 'alias claude="claude --model sonnet"' >> ~/.zshrc
source ~/.zshrc
# Verify the fix
claude --model sonnet -p "What model are you?" --max-turns 3

Step 3: Prevent

# Add model guard to shell profile
# ~/.zshrc
# Always default to Sonnet
alias claude='claude --model sonnet'
# Explicit Opus alias for intentional use (named to make it obvious)
alias claude-opus='claude --model opus'
# Print a reminder when using Opus
claude-opus() {
  echo "WARNING: Using Opus 4.6 ($15/$75 per MTok) -- 5x Sonnet cost"
  echo "Press Ctrl+C within 3 seconds to cancel..."
  sleep 3
  command claude --model opus "$@"
}

When Opus Is Actually Worth It

Not every Opus session is a mistake. Some tasks genuinely benefit from Opus 4.6’s superior reasoning, and spending 5x on tokens is justified by the quality improvement:

Justified Opus use cases:

  • Multi-system architecture design across 10+ files
  • Complex debugging where Sonnet failed after 2-3 attempts (the retry cost on Sonnet may exceed one Opus session)
  • Security audits requiring deep reasoning about attack vectors
  • Performance optimization requiring understanding of complex algorithmic tradeoffs

Never justified for Opus:

  • Simple edits (typo fixes, config changes)
  • Code reviews (reading comprehension, not deep reasoning)
  • Documentation generation
  • Test writing for straightforward functions

A useful rule of thumb: if the task would take a senior developer less than 15 minutes, Sonnet handles it fine. If it would take a senior developer an hour or more of deep thought, Opus may be worth the 5x premium.

# Decision helper: check task complexity before choosing model
# Quick tasks: always Sonnet
claude --model sonnet -p "Add error handling to the getUser function"
# Complex tasks: evaluate whether Opus is justified
# Sonnet attempt first, Opus only if Sonnet fails
claude --model sonnet --max-turns 15 -p "Debug the race condition in the payment processing pipeline"
# If this fails after 15 turns, then:
claude --model opus --max-turns 20 -p "Debug the race condition in the payment processing pipeline. Previous Sonnet attempt identified the issue is in src/services/payment.ts but could not determine root cause."

Cost Recovery

The overspent tokens cannot be refunded, but the financial impact can be quantified and future waste prevented:

Calculate overspend:
  Total Opus spend (from ccusage): $___
  Same tokens on Sonnet: Opus_spend / 5 = $___
  Overspend: Opus_spend x 0.8 = $___
Example:
  Opus sessions total: $25.00
  Same on Sonnet: $5.00
  Overspend: $20.00

Prevention saves: if the default was Opus for 1 week at moderate usage (500K tokens/day), the overspend is approximately $48.00. Fixing the alias takes 2 minutes and prevents all future incidents.

Prevention Rules for CLAUDE.md

## Model Selection
- Default model: Sonnet 4.6 (--model sonnet) for all standard tasks
- Opus 4.6: use ONLY for complex architecture decisions or multi-system debugging
- Before starting any Opus session, verify the task justifies 5x cost
- Check model in session header before proceeding with expensive tasks
- If session cost exceeds expectations, check model immediately

Team-Wide Model Governance

For teams, accidental Opus usage by one developer can blow the monthly budget. Implement governance:

# scripts/model-policy-check.sh
# Run in CI or as a pre-commit hook to verify model configuration
set -uo pipefail
echo "=== Model Policy Check ==="
# Check for Opus references in shell configs
for file in ~/.zshrc ~/.bashrc ~/.zprofile; do
  if [ -f "$file" ] && grep -qi "opus" "$file"; then
    echo "WARN: $file contains 'opus' reference -- verify it includes explicit --model flag"
  fi
done
# Check project settings for model specification
if [ -f ".claude/settings.json" ] && grep -qi "opus" ".claude/settings.json"; then
  echo "WARN: Project settings reference Opus -- verify this is intentional"
fi
# Check environment variables
if [ -n "${ANTHROPIC_MODEL:-}" ] && echo "$ANTHROPIC_MODEL" | grep -qi "opus"; then
  echo "ALERT: ANTHROPIC_MODEL is set to an Opus model: $ANTHROPIC_MODEL"
  echo "This overrides all session-level --model flags"
fi
echo ""
echo "Recommendation: alias claude='claude --model sonnet' in all developer shell profiles"

Add this check to the team onboarding process. A single developer accidentally using Opus for a full day can cost $50+ in API charges ($40 more than the same usage on Sonnet). Team-wide aliases and environment variable auditing prevent this. The 2-minute effort to set up the alias is the highest-ROI cost prevention measure available – it eliminates the most common cause of $20+ sessions.

Which model? → Take the 5-question quiz in our Model Selector.

Try it: Paste your error into our Error Diagnostic for an instant fix.

See Also