Claude Sonnet 5 is a proprietary large language model developed by Anthropic. Designed as a direct upgrade to Claude Sonnet 4.6, it offers enhanced capabilities in agentic workflows, complex reasoning, and coding, while maintaining the cost-efficiency typical of the Sonnet class. The model delivers a performance level close to flagship Opus-class models, particularly in reasoning, tool use, and long-horizon tasks.
Key Capabilities and Behavior Changes
Unlike its predecessor where reasoning was off by default, Claude Sonnet 5 has adaptive thinking enabled by default. This means that requests will consume reasoning tokens unless thinking is explicitly disabled. Rather than shipping separate model variants, Anthropic provides a single model (claude-sonnet-5) that supports two primary execution modes:
- Non-Reasoning Mode: Thinking can be completely turned off by setting the parameter
thinking: {type: "disabled"}. Disabling reasoning is recommended for straightforward, latency-sensitive tasks like classification, data formatting, and extraction, as it avoids the latency and token spend of reasoning blocks. - Reasoning Mode: When thinking is enabled, developers can control the depth of reasoning using the
effortparameter. The available effort levels are:- low: The most efficient setting for simpler or latency-sensitive tasks that still benefit from light reasoning.
- medium: A balanced level that provides moderate reasoning depth.
- high: The default effort setting, matching the default configuration of previous versions.
- xhigh: An extra-effort mode designed for complex coding and agentic tasks.
- max: The highest-capability mode, removing constraints on token spending to maximize performance.
Other changes in model behavior include strict enforcement of default sampling parameters; attempting to set non-default values for temperature, top_p, or top_k returns an API error. The model relies on system-prompt instructions rather than traditional sampling adjustments to guide its output style and behavior.
The model operates with a default and maximum context window of 1 million tokens and supports up to 128,000 output tokens. It utilizes an updated tokenizer that yields approximately 30% more tokens for identical input text compared to previous generations, improving overall performance across diverse text processing and retrieval benchmarks.
Architecture and Safety Safeguards
Claude Sonnet 5 is built on a transformer-based architecture, optimized for autonomous tasks like planning, browser usage, and terminal execution. Safety evaluations indicate that the model exhibits a lower rate of harmful behaviors compared to Claude Sonnet 4.6. It is equipped with real-time cybersecurity safeguards designed to detect and decline requests involving high-risk dual-use activities, resulting in a model that is safer for deployment in agentic scenarios.