Optimizing Claude 3.5 Sonnet for Production Workflows

Performance and Behavioral Shifts

Claude 3.5 Sonnet represents a shift in how Anthropic models handle long-running tasks. While previous iterations like Claude 3.5 Sonnet (referred to as Fable 5) often struggled to complete complex sequences, the 3.5 update finishes tasks more reliably. However, the model exhibits a higher propensity to hallucinate when it lacks an answer, reaching a 72.6% hallucination rate compared to 63.6% in the previous version. Additionally, the model is prone to triggering safety guardrails that may silently swap the active model to a weaker version without notifying the user, which can degrade output quality mid-session.

Optimization Techniques

To maximize efficiency and output quality, users should apply the following configuration and prompting strategies:

• Lower the Effort Setting: Testing by CodeRabbit shows that the "low" effort setting catches 61% of code issues compared to 57.1% on "high" effort, while completing tasks approximately 3 minutes faster.
• Audit Legacy Instructions: Remove outdated system instructions or prompt overrides that were designed to fix behaviors in older models. Use the claude-api-prompt-audit command in Claude Code to identify and remove redundant instructions.
• Enable Unattended Execution: Explicitly instruct the model that "nobody is watching" and that it has permission to proceed with any step that can be undone. This prevents the model from pausing to request confirmation for routine actions.
• Restrict Scope and Editing: Prevent the model from rewriting entire files when only a small change is required. Instruct the model to "edit only the specific lines needed" and explicitly define files or test suites that must be left untouched.
• Standardize Prose: To mitigate "mannered prose" and improve readability, instruct the model to "remove all mannered prose" and use direct language. This also helps break up long, dense paragraphs.
• Batch Complex Tasks: Avoid breaking features into granular, single-step prompts. The model performs significantly better when given the entire feature scope at once, leveraging its ability to maintain context over long sessions.

Before / After

• Hard Task Performance: On the most difficult client tasks, Claude 3.5 Sonnet finished approximately 40% faster and cost 58% less than the previous version.
• Token Efficiency: While cached reads are 75% cheaper, the model generates roughly 1.7 times more output tokens than its predecessor, which can lead to a 20% increase in total cost per task if not managed through scope restriction.