Last Updated: 29 July 2025
Anthropic recently unveiled Citations, a new feature for its Claude API models designed to improve the accuracy and trustworthiness of AI-generated responses.
This feature allows AI to “ground” its answers in source documents by providing detailed references to exact sentences and passages used to generate responses.
By offering verifiable, traceable outputs, Citations addresses one of the major challenges in AI applications: source reliability.
Anthropic’s Citations builds on Claude’s inherent focus on trustworthiness by enabling the AI to automatically cite claims inferred from source documents. This functionality is particularly beneficial for applications such as document summarization, complex Q&A, and customer support. With this feature, users can add source files like PDFs or plain text, and Claude automatically generates responses linked to specific sections of the documents. According to Anthropic, this approach has increased recall accuracy by up to 15% compared to most custom implementations.
Document summarization: Citations enables concise summaries of lengthy documents, like legal case files, with references to original sources for each key point.
Complex Q&A: It facilitates detailed answers derived from extensive corpora, such as financial statements, while linking responses back to their sources.
Customer support: By referencing multiple resources, such as product manuals or FAQs, it ensures accurate, traceable support solutions.
Citations integrates seamlessly with Anthropic’s Messages API and supports flexible workflows by processing source documents in chunks. This eliminates the need for extensive prompt engineering while maintaining a token-based pricing model. For example, processing a 100-page document costs approximately $0.30 with Claude 3.5 Sonnet or $0.08 with Claude 3.5 Haiku, making it a cost-effective solution for developers looking to reduce hallucinations and enhance output reliability.
Thomson Reuters: Using Citations to Power Legal AI Tools Thomson Reuters leverages Claude and its Citations feature for their AI platform, CoCounsel. According to Jake Heller, Head of Product at CoCounsel, the feature’s ability to accurately cite primary sources has improved trust in AI-generated content.
It has simplified maintaining their systems and enabled more reliable AI assistance for legal professionals.
Endex: Improving Financial Research Accuracy Endex, a provider of AI solutions for financial firms, reported significant improvements after implementing Citations. They observed a reduction in source hallucinations from 10% to 0% and a 20% increase in references per response.
CEO Tarun Amasa highlighted that the feature eliminated the need for complex prompt engineering and enhanced multi-stage financial research accuracy.
Anthropic’s Citations announcement comes amidst OpenAI’s launch of Operator, an AI agent capable of interacting with the web.
Unlike Citations, which focuses on source-based accountability, Operator is designed for broader web-based tasks, such as clicking and scrolling through GUIs.
Operator’s advanced reasoning allows it to self-correct and request user approval for sensitive tasks. However, Operator is currently limited to the US, available only to ChatGPT Pro subscribers, and struggles with complex interfaces.
Citations are available for Anthropic’s Claude 3.5 Sonnet and Claude 3.5 Haiku, two of the latest models in the Claude family, through the Anthropic API and Google Cloud’s Vertex AI. By prioritizing accuracy and transparency, this feature across Claude’s different models sets a new standard for AI accountability, making it particularly appealing for industries requiring high trust levels, such as legal, financial, and customer support sectors.
