In March 2026, developers using Claude Code reported a decline in performance, with AMD's head of AI, Stella Laurenzo, identifying a 67% reduction in reasoning output across 6,852 GitHub session logs. Anthropic revealed that this performance dip was not caused by model degradation but by a March 4 update that lowered the default 'Effort' setting from High to Medium to reduce latency. This setting determines how much work the AI is willing to invest in a task, such as file reading and test verification, rather than its inherent reasoning capability. Anthropic restored the default setting to High on April 7 and reset usage quotas for paying subscribers.

The 'Effort' parameter dictates autonomous workload capacity; High Effort can generate up to 7x more tokens for task verification than Low Effort. Because of this, a smaller model on High Effort can outperform a larger model on Low Effort during complex multi-step tasks. Anthropic now categorizes its models by task-specific personas: Sonnet functions as a thorough all-rounder, Opus acts as an experienced specialist, and Fable is reserved for the most complex engineering challenges. This dynamic illustrates a shift in AI coding from raw model strength to agent orchestration, where developers manage AI agents like project managers.

The competitive landscape has intensified following OpenAI's July 9 release of GPT-5.6 and the ChatGPT Work feature, which compete directly with Claude Cowork. While real-world tests indicate that GPT-5.6 Sol achieves high-quality results comparable to Anthropic's tools—specifically in aesthetic-heavy web generation tasks—it faces challenges with rapid agent quota consumption and high costs. Conversely, Claude Cowork maintains superior maturity in local desktop workflows and document organization. Current industry benchmarks suggest that for production-oriented programs, Claude Cowork remains the more reliable choice for integrated tasks, while the core productivity skill for developers is learning to optimize model selection and effort levels.

In March 2026, developers using Claude Code reported a decline in performance, with AMD's head of AI, Stella Laurenzo, identifying a 67% reduction in reasoning output across 6,852 GitHub session logs. Anthropic revealed that this performance dip was not caused by model degradation but by a March 4 update that lowered the default 'Effort' setting from High to Medium to reduce latency. This setting determines how much work the AI is willing to invest in a task, such as file reading and test verification, rather than its inherent reasoning capability. Anthropic restored the default setting to High on April 7 and reset usage quotas for paying subscribers.

The 'Effort' parameter dictates autonomous workload capacity; High Effort can generate up to 7x more tokens for task verification than Low Effort. Because of this, a smaller model on High Effort can outperform a larger model on Low Effort during complex multi-step tasks. Anthropic now categorizes its models by task-specific personas: Sonnet functions as a thorough all-rounder, Opus acts as an experienced specialist, and Fable is reserved for the most complex engineering challenges. This dynamic illustrates a shift in AI coding from raw model strength to agent orchestration, where developers manage AI agents like project managers.

The competitive landscape has intensified following OpenAI's July 9 release of GPT-5.6 and the ChatGPT Work feature, which compete directly with Claude Cowork. While real-world tests indicate that GPT-5.6 Sol achieves high-quality results comparable to Anthropic's tools—specifically in aesthetic-heavy web generation tasks—it faces challenges with rapid agent quota consumption and high costs. Conversely, Claude Cowork maintains superior maturity in local desktop workflows and document organization. Current industry benchmarks suggest that for production-oriented programs, Claude Cowork remains the more reliable choice for integrated tasks, while the core productivity skill for developers is learning to optimize model selection and effort levels.