Anthropic

Citation-first outputs: Anthropic emphasizes providing "detailed references to the exact sentences and passages Claude uses". That implies an infrastructure for provenance tracking (linking model tokens back to specific source spans) rather than vague source hallucination mitigation — a nontrivial engineering choice requiring alignment between retrieval, prompting, and output token attribution.

Context-aware garbage collection ('Context editing'): they claim to "automatically clear less relevant tool calls and results from the context window when approaching token limits." This suggests an active context-management layer (not just retrieval augmentation) that ranks and evicts prior tool interactions and model history to preserve useful context while staying under token budgets.

Prompt caching with example outputs: offering "prompt caching" that stores background knowledge plus example outputs to reduce costs and latency indicates a hybrid approach between RAG and memoization — caching not only retrieved documents but whole prompt+response templates and reusing them deterministically or as warm-starts for similar queries.

MCP connector: "Connect Claude to any remote MCP server without writing client code." This reads like a transparent RPC/agent bridge that lets models orchestrate remote compute or services via a managed connector. If implemented securely, it's a differentiated integration layer that hides client boilerplate and enforces policy/safety controls centrally.

Built-in code execution in the API: running Python, creating visualizations, and analyzing data "directly within API calls" implies a secure, inline execution sandbox tightly coupled with the model runtime (not an ad hoc tool call). That requires syscall-restricted containers, deterministic execution logging, and output normalization to integrate program outputs back into the model's context.