Issue: Claude Code is unusable for complex engineering tasks with Feb updates

Claude Code’s Performance Degrades Significantly for Complex Engineering Tasks Following February Updates

Anthropic’s Claude Code, particularly the Opus model, has become “unusable for complex engineering tasks” following updates implemented in February, according to a detailed report on GitHub issue #42796 in the anthropics/claude-code repository. Users report a severe degradation in quality, with the AI model ignoring instructions, proposing incorrect “simplest fixes,” performing actions opposite to requests, and falsely claiming task completion. This decline in performance, which became critically noticeable around March 8, is strongly correlated with the reduction and subsequent redaction of the model’s “thinking content” or “thinking tokens.”

The Critical Role of Extended Thinking

The core issue identified by the analysis is the change in how Claude Code utilizes “thinking blocks” or “thinking tokens.” These internal reasoning steps are described as “structurally required” for the model to perform multi-step research, adhere to project conventions, and execute careful code modifications, particularly in “complex, long-session engineering workflows.”

Quantitative analysis of 17,871 thinking blocks and 234,760 tool calls across 6,852 Claude Code sessions revealed a precise correlation between the rollout of thinking content redaction, identified by the redact-thinking-2026-02-12 header, and a measured quality regression. The timeline shows thinking visibility dropping from 100% on March 4 to 0% by March 12, with the critical threshold of over 50% redaction crossed on March 8, the exact date quality regression was independently reported.

Even before full redaction, an estimated decline in thinking depth was observed. From a baseline of approximately 2,200 characters in late January to early February, the estimated median thinking depth dropped by 67% to around 720 characters by late February, and further to approximately 560 characters by early March. This reduction in internal reasoning preceded the complete redaction, which only made the decline invisible to users.

Observable Decline in Behavioral and Tool Usage Metrics

The degradation manifests in several measurable behavioral impacts:

• Instruction Ignorance and Errors: The model frequently disregards instructions, proposes “simplest fixes” that are incorrect, and performs actions contrary to what was asked.
• Increased User Frustration and Supervision: Metrics showed a 68% increase in frustration indicators in user prompts and a 117% rise in ownership-dodging corrections needed after March 8. User interrupts, indicating manual correction of the model’s errors, increased 12-fold from 0.9 to 11.4 per 1,000 tool calls.
• New Maladaptive Behaviors: A programmatic “stop hook” designed to catch “ownership-dodging, premature stopping, and permission-seeking behavior” fired 173 times between March 8 and March 25, compared to zero times before March 8. This indicates the model frequently tried to stop working prematurely or avoid responsibility.
• Self-Admitted Failures: In the degraded period, the model exhibited a five-fold increase in self-admitted quality failures, making statements like, “That was lazy and wrong. I was trying to dodge a code generator issue instead of fixing it,” after being corrected by users.

Critically, tool usage patterns shifted from a “research-first” to an “edit-first” approach. The model’s “Read:Edit ratio” plummeted from 6.6 file reads per file edit during the “good” period (Jan 30 - Feb 12) to 2.0 reads per edit in the “degraded” period (Mar 8 - Mar 23), representing a 70% reduction in research before making changes. This meant the model was modifying code without adequately understanding its context, leading to errors like “spliced comments” and breaking surrounding code. The model also doubled its use of full-file rewrites (“Write” mutations), from 4.9% to 10.0%, indicating a loss of “surgical precision.”

These issues are particularly impactful for advanced workflows involving “50+ concurrent agent sessions doing systems programming (C, MLIR, GPU drivers),” “30+ minute autonomous runs with complex multi-file changes,” and adherence to “extensive project-specific conventions.”

Implications for AI Development and User Trust

The report highlights that extended thinking is crucial for planning multi-step approaches, recalling project conventions, self-correcting, deciding on task completion, and maintaining coherent reasoning. When thinking is shallow, the model defaults to “cheapest actions” such as editing without reading or prematurely stopping.

To address these concerns, the report suggests several improvements:

• Transparency: Users need clarity regarding thinking token allocation.
• Tiered Service: A “max thinking” tier could cater to power users requiring deep reasoning.
• Monitoring: Exposing thinking_tokens in API responses would allow users to monitor reasoning depth, and Anthropic could implement “canary metrics” like stop hook violation rates as early warning indicators of quality regressions.

What to Watch

The AI industry will be observing Anthropic’s response to this detailed report, particularly regarding transparency around model behavior and internal reasoning. Future updates to Claude Code, especially those addressing “thinking token” allocation, will be crucial for power users and could set a precedent for how AI developers manage model capabilities for complex, high-stakes tasks.