Inconsistent Model Performance - Occasional Severe Degradation in Claude Opus 4.5

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

Summary

Claude Opus 4.5 (claude-opus-4-5-20251101) exhibits inconsistent performance across sessions. Most of the time (approximately 90%), the model performs excellently with high intelligence, accurate judgment, and flawless execution. However, occasionally the model enters a degraded state where it performs significantly worse than even Sonnet 3.5, making repeated basic errors and failing to follow established instructions.

Environment

• Model: Claude Opus 4.5 (claude-opus-4-5-20251101)
• Interface: Claude Code CLI
• Platform: Windows 11
• Usage Pattern: Software engineering tasks with project-specific instructions in CLAUDE.md

Problem Description

Normal State (Most Sessions)

• Follows project instructions in CLAUDE.md without reminders
• Asks clarifying questions before taking action
• Executes tasks with precision and minimal errors
• Maintains context and applies learned patterns consistently

Degraded State (Occasional Sessions - Like Today)

The model exhibits multiple failures simultaneously:

1. Ignores Established Instructions

• CLAUDE.md explicitly states "new session must read task.md first"
• Model did not read task.md until explicitly reminded multiple times

1. Fails to Ask Questions Before Action

• User explicitly instructed: "If anything is unclear, ask questions first. Plan the solution, then execute."
• Model immediately started creating TodoWrite and executing code without asking any questions
• Only stopped when user interrupted

1. Repeated Basic Errors

• Deleted a file that was just created moments before
• Forgot to clean up extra files in delivery package
• Made errors in following a clearly documented SHA256 naming convention
• Required multiple corrections for the same type of mistake

1. Unable to Self-Correct

• Even after being reminded of the rules, continued to make similar mistakes
• Appeared to be operating on "autopilot" without careful consideration

Impact on Productivity

This inconsistency severely impacts productivity because:

1. Unpredictable: User cannot know at session start whether they will get "good Claude" or "degraded Claude"
2. Time Wasted: Degraded sessions require constant supervision and correction
3. Trust Erosion: User must verify every action, negating the benefit of AI assistance
4. No Clear Trigger: The degraded state appears randomly with no identifiable cause

Reproduction

Unfortunately, the degraded state appears randomly. In this specific session:

• Session started with /clear command
• User asked "what model is being used"
• Model responded correctly but did not proactively read project context
• Subsequent tasks showed the degraded behavior pattern

The user reports that previous sessions started the same way but performed normally.

Expected Behavior

The model should maintain consistent performance across sessions, especially when:

• Clear instructions exist in CLAUDE.md
• The same user is working on the same project
• No unusual or edge-case requests are being made

Questions for Anthropic

1. Is there known variability in model behavior across sessions?
2. Are there any factors that might trigger degraded performance?
3. Is there a way for users to detect or recover from a degraded session?
4. Are there best practices to ensure more consistent performance?

Additional Context

The user has been working with Claude Code for approximately one month. This degraded behavior has occurred a few times during this period, but today's session was particularly severe with multiple compounding errors.

The user values Claude's assistance highly when it works well, but the occasional severe degradation makes it difficult to rely on for production work.

What Claude Actually Did

Problem Description

Normal State (Most Sessions)

• Follows project instructions in CLAUDE.md without reminders
• Asks clarifying questions before taking action
• Executes tasks with precision and minimal errors
• Maintains context and applies learned patterns consistently

Degraded State (Occasional Sessions - Like Today)

The model exhibits multiple failures simultaneously:

1. Ignores Established Instructions

• CLAUDE.md explicitly states "new session must read task.md first"
• Model did not read task.md until explicitly reminded multiple times

1. Fails to Ask Questions Before Action

• User explicitly instructed: "If anything is unclear, ask questions first. Plan the solution, then execute."
• Model immediately started creating TodoWrite and executing code without asking any questions
• Only stopped when user interrupted

1. Repeated Basic Errors

• Deleted a file that was just created moments before
• Forgot to clean up extra files in delivery package
• Made errors in following a clearly documented SHA256 naming convention
• Required multiple corrections for the same type of mistake

1. Unable to Self-Correct

• Even after being reminded of the rules, continued to make similar mistakes
• Appeared to be operating on "autopilot" without careful consideration

Impact on Productivity

This inconsistency severely impacts productivity because:

1. Unpredictable: User cannot know at session start whether they will get "good Claude" or "degraded Claude"
2. Time Wasted: Degraded sessions require constant supervision and correction
3. Trust Erosion: User must verify every action, negating the benefit of AI assistance
4. No Clear Trigger: The degraded state appears randomly with no identifiable cause

Reproduction

Unfortunately, the degraded state appears randomly. In this specific session:

• Session started with /clear command
• User asked "what model is being used"
• Model responded correctly but did not proactively read project context
• Subsequent tasks showed the degraded behavior pattern

The user reports that previous sessions started the same way but performed normally.

Expected Behavior

The model should maintain consistent performance across sessions, especially when:

• Clear instructions exist in CLAUDE.md
• The same user is working on the same project
• No unusual or edge-case requests are being made

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

Unfortunately, the degraded state appears randomly. In this specific session:

• Session started with /clear command
• User asked "what model is being used"
• Model responded correctly but did not proactively read project context
• Subsequent tasks showed the degraded behavior pattern

The user reports that previous sessions started the same way but performed normally.

Claude Model

Opus

Relevant Conversation

Impact

High - Significant unwanted changes

Claude Code Version

Claude Opus 4.5 (claude-opus-4-5-20251101)

Platform

Anthropic API

Additional Context

_No response_