Claude Sonnet 4

Claude Sonnet 4.6 implements a hybrid reasoning architecture where users can explicitly trigger extended thinking mode to enable step-by-step problem decomposition before generating responses. The model performs internal chain-of-thought reasoning (hidden from users) and can be configured with fine-grained thinking effort levels via API parameters, trading off latency and cost for reasoning depth. This differs from standard token-by-token generation by allocating compute budget to pre-response deliberation rather than streaming output.

Claude Sonnet 4.6 achieves 'frontier coding performance' through transformer-based understanding of code structure, context, and intent across multiple files. The model can analyze entire codebases (up to 1M context window in beta), generate code that respects existing patterns and dependencies, and perform refactoring operations that maintain semantic correctness. Implementation leverages the full context window to maintain awareness of imports, type definitions, and architectural constraints without requiring explicit AST parsing or language-specific plugins.

Claude Sonnet 4.6 offers Claude Managed Agents, a separate infrastructure from the standard Messages API that provides fully managed agent hosting with stateful sessions and persistent event history. Developers define agent behavior via a configuration file (tools, instructions, model), and Anthropic manages session state, tool invocation, and error handling. This differs from the Messages API by providing built-in session management and persistent memory without requiring developers to implement state management logic.

Claude Sonnet 4.6 supports understanding and generation in multiple languages, enabling translation, multilingual content analysis, and cross-language reasoning. The model can process input in one language and generate output in another, or analyze multilingual documents and extract information across language boundaries. Implementation leverages the transformer's multilingual training to handle language mixing and code-switching without explicit language detection or separate translation models.

Claude Sonnet 4.6 includes built-in safety features to reduce harmful outputs, including guardrails for hallucination reduction, jailbreak mitigation, and content filtering. These are implemented at the model level (training-time alignment) and optionally at the API level (request-time filtering). Developers can configure safety settings per-request, and Anthropic provides documentation on responsible use patterns. The model refuses harmful requests and explains why, rather than generating harmful content.

Claude Sonnet 4.6 supports context editing capabilities that allow developers to modify conversation history, remove messages, or adjust context mid-conversation without restarting. This is implemented via API parameters that allow selective message deletion or replacement, enabling dynamic conversation management. Developers can use context editing to remove sensitive information, correct errors, or optimize token usage by removing less relevant messages.

Claude Sonnet 4.6 provides a token counting API that allows developers to estimate costs before making API requests. The count_tokens endpoint accepts text, images, and tool definitions and returns the exact token count that would be billed. This enables budget forecasting, cost optimization, and request planning without making actual API calls. Token counting is implemented as a separate, low-cost API endpoint (typically free or minimal cost).

Claude Sonnet 4.6 can analyze screenshots and execute browser/desktop automation tasks by understanding visual layouts, identifying UI elements, and generating appropriate actions (clicks, text input, navigation). The model receives image input of the current screen state, reasons about the task, and outputs structured commands (via built-in computer-use tool) to interact with the GUI. This enables autonomous task execution in digital environments without requiring explicit element selectors or DOM access.

Claude Sonnet 4.6 includes a built-in web search tool that allows the model to query the internet and retrieve current information during conversation. When enabled, the model can autonomously decide when to search, fetch web content, and synthesize results into responses. This is implemented as a native tool in the Messages API (alongside code execution and computer use), allowing developers to enable/disable web search per-request without additional API calls or external search service integration.

Claude Sonnet 4.6 supports structured output mode where developers define a JSON schema and the model is constrained to generate responses matching that schema exactly. This is implemented via the Messages API's response_format parameter with json_schema specification, ensuring outputs are valid JSON that can be directly parsed and used in downstream systems without additional validation or parsing logic. The model's token generation is constrained to only produce valid schema-compliant outputs.

Claude Sonnet 4.6 implements prompt caching where frequently-used context (system prompts, documents, code files, instructions) is cached server-side after the first request. Subsequent requests reusing the same cached context are charged at 90% discount on cached tokens (input tokens cost $0.30/M instead of $3/M for Sonnet 4.6). Caching is transparent to developers — any repeated input tokens are automatically cached and reused without explicit cache management.

Claude Sonnet 4.6 offers a Batch API that processes multiple requests asynchronously in a single batch, providing up to 50% cost reduction compared to standard API calls. Developers submit a batch of requests (JSONL format) and receive results after processing completes (typically within 24 hours). Batch processing is implemented as a separate API endpoint with different pricing and SLA, allowing developers to trade latency for cost on non-urgent workloads.

Claude Sonnet 4.6 can analyze images (PNG, JPEG, GIF, WebP formats) to extract information, answer questions about visual content, and perform OCR-like text extraction. The model receives image input as base64-encoded data or URLs and generates text descriptions, answers to visual questions, or structured data extracted from images. Vision capability is integrated into the standard Messages API without separate endpoints, allowing seamless mixing of text and image inputs in conversations.

Claude Sonnet 4.6 can execute Python code in a sandboxed environment, allowing the model to run scripts, perform calculations, and validate outputs. Code execution is implemented as a built-in tool in the Messages API (alongside web search and computer use), enabling the model to autonomously write and run code to solve problems or verify results. The sandbox provides a Python runtime with common libraries (NumPy, Pandas, Matplotlib, etc.) and file system access within the sandbox.

Claude Sonnet 4.6 supports parallel tool use where the model can invoke multiple tools (web search, code execution, computer use, etc.) simultaneously in a single response, rather than sequentially. This is implemented via the Messages API's tool_use capability, allowing the model to parallelize independent operations (e.g., search multiple queries, execute multiple code blocks) and combine results. Developers can also enable strict tool use mode to enforce that responses only contain tool calls without additional text.