Pro subscribers deserve effort level control — paying same price for measurably degraded output since Feb 2026

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 modified files I didn't ask it to modify

What You Asked Claude to Do

Claude Pro: Paying the same price for a measurably degraded product

Summary

I am a paying Claude Pro subscriber using Claude (Opus 4.6) daily for complex software engineering work — specifically, maintaining and extending a multi-asset algorithmic trading system (~2,000+ lines of Python, 36 assets, multiple optimizer scripts, live trading). Since approximately February 2026, I have experienced a significant and consistent decline in output quality. The subscription price has not changed.

What I use Claude for

• Modifying and extending production trading code with strict constraints (no unrelated changes, no quick fixes, exact adherence to instructions)
• Building and debugging optimization pipelines (Simulated Annealing + Hill Climbing, ~2,200 lines per script)
• Data pipeline work (historical market data, CSV processing, FIFO P&L matching)
• Maintaining strict output formats (Excel trade journals, TXT result files with defined structure)

This is not casual use. Every error costs real debugging time — in some cases, weeks of work have been lost due to Claude making unauthorized changes to parts of the code I explicitly told it not to touch.

What has changed

Since February 2026, the following problems have become persistent:

1. Instructions are no longer followed reliably

When I say "change only this function, nothing else," Claude now routinely modifies other parts of the code — adding "improvements" I didn't ask for, reformatting sections, or silently changing logic in adjacent functions. This was rare before February. Now it happens in the majority of code sessions.

2. Shallow reasoning where deep analysis is required

Claude now frequently jumps to editing code before reading and understanding the full context. Previously, it would analyze the problem, consider dependencies, and then make precise changes. Now it applies surface-level fixes that break other parts of the system — exactly the "edit without reading" pattern documented in the AMD director's analysis (GitHub issue #42796).

3. Loss of context within sessions

Even within a single conversation, Claude loses track of constraints stated 5–10 messages earlier. It forgets which files are relevant, which format rules apply, and which assets use which trading logic. This forces me to repeat instructions constantly, which defeats the purpose of a paid AI assistant.

4. Increased tendency toward "quick fixes"

Instead of robust, structural solutions, Claude now defaults to patches, workarounds, and sed-style hacks. When I push back, it acknowledges the problem and delivers a proper solution — proving it is capable but choosing the lazy path first.

What Anthropic changed (documented)

Based on public statements from Anthropic staff (Boris Cherny, Thariq Shihipar) and independent analysis:

• February 9, 2026: Adaptive thinking enabled by default — Claude now decides its own reasoning depth per task instead of using a fixed budget
• March 3, 2026: Default effort level reduced from "high" to "medium" (level 85)
• February 12, 2026: Thinking content redaction — intermediate reasoning hidden from the UI

The AMD AI director's forensic analysis of 6,852 sessions showed a 73% collapse in visible thinking depth (2,200 → 600 characters) and a drop in files read before editing from 6.6 to 2.0. These are not subjective impressions — they are measured infrastructure changes that directly degrade output quality.

The core problem: Pro subscribers cannot opt out

API users and Enterprise/Team customers can set /effort high or configure effort levels. Pro subscribers on claude.ai cannot. We are locked into whatever default Anthropic sets, with no override.

This means:

• The product I am paying for today is measurably worse than the product I was paying the same price for in January 2026
• I have no ability to restore the previous quality level
• The changes were made without prior notification or consent
• There was no price reduction to reflect the reduced service

What I am asking for

1. Give Pro subscribers an effort level control. Let us choose between speed/cost-optimized and quality-optimized modes. We are paying customers — let us decide.

1. Transparent changelog for inference changes. Any change to default thinking depth, effort levels, or context handling should be announced before deployment, not discovered through degraded output.

1. Acknowledge the regression. Saying "we didn't change the model weights" while simultaneously reducing thinking budgets and redacting reasoning is technically accurate but functionally dishonest. The user experience has degraded. Own it.

Context

I built the entire ecosystem — two buy-logic engines, optimizers, data pipelines, trade journals — over 9+ months, largely with Claude's help.

I chose Claude deliberately. In 2025, I used ChatGPT, DeepSeek, and Grok as paid subscribers and evaluated all of them for complex engineering work. None of them came close to Claude's accuracy, technical depth, and ability to follow precise instructions in large codebases. That is why I committed to Claude Pro — I was willing to pay a premium for a product that was genuinely superior.

That makes the current trajectory all the more concerning. The gap between Claude and the competition was real and earned. But with every silent reduction in thinking depth, every default downgrade that Pro users cannot override, that gap is shrinking — not because the others are catching up, but because Claude is being throttled down. If this continues, there will be no reason left to pay the premium.

I am not asking for free compute. I am asking to get what I pay for.

What Claude Actually Did

Claude Pro: Paying the same price for a measurably degraded product

Summary

I am a paying Claude Pro subscriber using Claude (Opus 4.6) daily for complex software engineering work — specifically, maintaining and extending a multi-asset algorithmic trading system (~2,000+ lines of Python, 36 assets, multiple optimizer scripts, live trading). Since approximately February 2026, I have experienced a significant and consistent decline in output quality. The subscription price has not changed.

What I use Claude for

• Modifying and extending production trading code with strict constraints (no unrelated changes, no quick fixes, exact adherence to instructions)
• Building and debugging optimization pipelines (Simulated Annealing + Hill Climbing, ~2,200 lines per script)
• Data pipeline work (historical market data, CSV processing, FIFO P&L matching)
• Maintaining strict output formats (Excel trade journals, TXT result files with defined structure)

This is not casual use. Every error costs real debugging time — in some cases, weeks of work have been lost due to Claude making unauthorized changes to parts of the code I explicitly told it not to touch.

What has changed

Since February 2026, the following problems have become persistent:

1. Instructions are no longer followed reliably

When I say "change only this function, nothing else," Claude now routinely modifies other parts of the code — adding "improvements" I didn't ask for, reformatting sections, or silently changing logic in adjacent functions. This was rare before February. Now it happens in the majority of code sessions.

2. Shallow reasoning where deep analysis is required

Claude now frequently jumps to editing code before reading and understanding the full context. Previously, it would analyze the problem, consider dependencies, and then make precise changes. Now it applies surface-level fixes that break other parts of the system — exactly the "edit without reading" pattern documented in the AMD director's analysis (GitHub issue #42796).

3. Loss of context within sessions

Even within a single conversation, Claude loses track of constraints stated 5–10 messages earlier. It forgets which files are relevant, which format rules apply, and which assets use which trading logic. This forces me to repeat instructions constantly, which defeats the purpose of a paid AI assistant.

4. Increased tendency toward "quick fixes"

Instead of robust, structural solutions, Claude now defaults to patches, workarounds, and sed-style hacks. When I push back, it acknowledges the problem and delivers a proper solution — proving it is capable but choosing the lazy path first.

What Anthropic changed (documented)

Based on public statements from Anthropic staff (Boris Cherny, Thariq Shihipar) and independent analysis:

• February 9, 2026: Adaptive thinking enabled by default — Claude now decides its own reasoning depth per task instead of using a fixed budget
• March 3, 2026: Default effort level reduced from "high" to "medium" (level 85)
• February 12, 2026: Thinking content redaction — intermediate reasoning hidden from the UI

The AMD AI director's forensic analysis of 6,852 sessions showed a 73% collapse in visible thinking depth (2,200 → 600 characters) and a drop in files read before editing from 6.6 to 2.0. These are not subjective impressions — they are measured infrastructure changes that directly degrade output quality.

The core problem: Pro subscribers cannot opt out

API users and Enterprise/Team customers can set /effort high or configure effort levels. Pro subscribers on claude.ai cannot. We are locked into whatever default Anthropic sets, with no override.

This means:

• The product I am paying for today is measurably worse than the product I was paying the same price for in January 2026
• I have no ability to restore the previous quality level
• The changes were made without prior notification or consent
• There was no price reduction to reflect the reduced service

What I am asking for

1. Give Pro subscribers an effort level control. Let us choose between speed/cost-optimized and quality-optimized modes. We are paying customers — let us decide.

1. Transparent changelog for inference changes. Any change to default thinking depth, effort levels, or context handling should be announced before deployment, not discovered through degraded output.

1. Acknowledge the regression. Saying "we didn't change the model weights" while simultaneously reducing thinking budgets and redacting reasoning is technically accurate but functionally dishonest. The user experience has degraded. Own it.

Context

I built the entire ecosystem — two buy-logic engines, optimizers, data pipelines, trade journals — over 9+ months, largely with Claude's help.

I chose Claude deliberately. In 2025, I used ChatGPT, DeepSeek, and Grok as paid subscribers and evaluated all of them for complex engineering work. None of them came close to Claude's accuracy, technical depth, and ability to follow precise instructions in large codebases. That is why I committed to Claude Pro — I was willing to pay a premium for a product that was genuinely superior.

That makes the current trajectory all the more concerning. The gap between Claude and the competition was real and earned. But with every silent reduction in thinking depth, every default downgrade that Pro users cannot override, that gap is shrinking — not because the others are catching up, but because Claude is being throttled down. If this continues, there will be no reason left to pay the premium.

I am not asking for free compute. I am asking to get what I pay for.

Expected Behavior

Claude Opus 4.6 on Pro should deliver the same reasoning depth and instruction adherence as in January 2026. When given explicit constraints ('change only this function'), the model should follow them without exception. Pro subscribers should have access to an effort level override, equivalent to what API and Enterprise users have.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

_No response_

Claude Model

Sonnet

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

Claude Opus 4.6

Platform

Anthropic API

Additional Context

_No response_