Anthropic: Claude Sonnet 4

Claude Sonnet 4 maintains coherent multi-turn conversations with up to 200K token context window, using transformer-based attention mechanisms to track conversation history and reference previous exchanges. The model employs constitutional AI training to ensure consistent reasoning across long conversations while managing context efficiently through selective attention patterns rather than naive concatenation.

Claude Sonnet 4 generates production-ready code across 40+ programming languages using transformer-based code understanding trained on vast open-source repositories and SWE-bench datasets. The model applies structural awareness through implicit AST-like reasoning patterns, enabling it to generate contextually appropriate code that respects language idioms, type systems, and existing codebase patterns without explicit tree-sitter parsing.

Claude Sonnet 4 processes images (JPEG, PNG, WebP, GIF formats) up to 20MB through a vision transformer backbone, extracting text via OCR, identifying objects, analyzing layouts, and reasoning about visual content. The model integrates vision and language understanding through a unified transformer architecture, allowing it to answer questions about images, describe scenes, and extract structured data from visual documents without separate API calls.

Claude Sonnet 4 generates structured outputs conforming to user-specified JSON schemas through constrained decoding, where the model's token generation is restricted to valid JSON paths that satisfy the schema constraints. This approach uses a constraint-aware sampling algorithm that prevents invalid outputs at generation time rather than post-processing, ensuring 100% schema compliance without requiring output validation or retry logic.

Claude Sonnet 4 supports tool calling through a native function-calling API where developers define tools as JSON schemas and the model decides when to invoke them, returning structured tool-use blocks with arguments. The implementation uses a separate token stream for tool decisions, allowing the model to reason about which tools to use before committing to a function call, and supports parallel tool invocation (multiple tools in a single response) for efficient orchestration.

Claude Sonnet 4 implements prompt caching where frequently-used context (system prompts, documents, code files) is cached server-side after the first request, reducing token processing cost by 90% and latency by 50-70% on subsequent requests with identical cached content. The caching uses a content-hash based key system that automatically detects when cached content can be reused, requiring no explicit cache management from developers.

Claude Sonnet 4 offers a batch processing API that accepts multiple requests in a single JSONL file, processes them asynchronously with 50% cost reduction compared to standard API calls, and returns results in a separate output file. The batch system uses off-peak compute resources and optimizes token utilization across requests, trading latency (12-24 hour turnaround) for significant cost savings, making it ideal for non-time-sensitive workloads.

Claude Sonnet 4 is trained using Constitutional AI (CAI), where a set of principles (constitution) guides model behavior during training and inference. The model learns to self-critique and revise outputs to align with these principles, reducing harmful outputs and improving factuality. While the base constitution is fixed, developers can influence behavior through system prompts that specify values, constraints, or guidelines, effectively creating application-specific alignment without model retraining.

Claude Sonnet 4 supports extended thinking mode where the model allocates additional compute to reasoning before generating a response, using an internal chain-of-thought process that explores multiple solution paths and validates reasoning before committing to an answer. This approach increases latency by 2-5x but significantly improves accuracy on complex tasks like mathematical proofs, multi-step logic puzzles, and intricate code debugging by enabling deeper exploration of the problem space.