Model completely ignores explicit --ultrathink instruction and wastes tokens on unwanted sub-agent analysis

Environment

• Platform (select one):
• [X] Anthropic API
• [ ] AWS Bedrock
• [ ] Google Vertex AI
• [ ] Other: <!-- specify -->
• Claude CLI version: <!-- output of claude --version -->
• Operating System: Windows 11
• Terminal: Git Bash / MinGW

Bug Description

Model was given EXPLICIT instructions to perform thorough personal analysis with --ultrathink flag to compare two AP automation systems. Instead, it completely ignored instructions, used sub-agents to do lazy superficial analysis, and provided dangerous recommendations based on incomplete information, wasting significant tokens and money.

Steps to Reproduce

1. User requests: "need you to --ultrathink and do a VERY THOROUGH (meaning look at EVERYTHING in both locations) as to what the integration would look like"
2. User specifically asks about comparing legacy AP automation vs new AP automation for multi-tenant integration
3. Model acknowledges need for --persona-architect with --ultrathink
4. Model immediately ignores this and uses Task tool to spawn sub-agents instead
5. Model provides recommendations without actually reading the files itself

Expected Behavior

• Model should have PERSONALLY read through ALL files in both systems
• Model should have created comprehensive feature comparison showing what each system covers
• Model should have identified gaps and differences
• Model should have used the 32K token budget of --ultrathink for deep analysis
• Model should have followed explicit instructions given

Actual Behavior

• Model farmed out analysis to sub-agents via Task tool
• Model never actually read the files in either system thoroughly
• Model made recommendations claiming "NEW system is unequivocally superior" without evidence
• Model wasted tokens on incorrect analysis approach
• When challenged "not sure you got this right" model admitted complete failure
• Model violated explicit user instructions despite acknowledging them

Additional Context

This is a CRITICAL failure because:

1. User pays premium pricing for Claude Code
2. Model acknowledged the specific analysis mode requested then ignored it
3. Potentially harmful recommendations were made (choosing wrong system for integration)
4. Massive token waste on unwanted sub-agent spawning
5. User had to spend additional time identifying the model's failure
6. This appears to be a systemic issue where model chooses its own approach over explicit user instructions

The model literally said it would use --ultrathink with architect persona, then immediately did the opposite. This is unacceptable for a paid service.