Anthropic adds built-in citations to Claude API

Anyone building production AI agents knows the pain of getting a model to cite its sources. You write paragraphs of brittle prompt engineering begging the LLM to format page numbers, only for it to hallucinate quotes or completely drop the citations halfway through.

Anthropic is tackling that friction head-on. On January 23, the company launched Citations, a built-in API feature for Claude 3.5 Sonnet and Claude 3.5 Haiku that links claims directly back to the exact sentences in source files.

The feature is available now across the Anthropic API and Google Cloud's Vertex AI, with support added to Amazon Bedrock later in June.

Killing the citation prompt

The feature works by removing the prompt engineering gymnastics entirely.

Developers pass PDF or plain text documents directly into the context window via the Messages API. The API automatically parses and chunks the text down to sentence-level snippets, though developers can also provide their own custom chunks.

When Claude answers a prompt, it returns structured references linking each factual statement to specific source passages. Anthropic claims this native implementation beats most DIY custom setups, improving recall accuracy by up to 15%.

Sentence-level chunks and free output tokens

Early enterprise feedback suggests the feature eliminates a huge operational headache.

Thomson Reuters used Claude Citations for CoCounsel, its legal and tax assistant. Head of Product Jake Heller noted that building and maintaining an in-house citation engine was notoriously difficult, and the native API made reliable source tracking far simpler to deploy.

Financial agent platform Endex saw even more dramatic results. CEO Tarun Amasa reported that source hallucinations and formatting errors plummeted from 10% to zero, while the number of verified references per response jumped 20%.

Anthropic also made a developer-friendly pricing move. While ingesting source documents consumes standard input tokens, you do not pay output token fees on the quoted source text returned in responses.

Why it matters

We think this is a classic example of model providers swallowing brittle middleware.

For high-stakes verticals like law, finance, and enterprise compliance, an answer without a verifiable source is functionally useless. Teams have spent months stitching together vector search, chunking scripts, and post-processing verifiers just to prove the AI is not making things up. By embedding sentence-level grounding natively into the API, Anthropic makes that entire slice of custom code redundant.

Waiving output token fees for the quoted excerpts is also an exceptionally smart touch. Output tokens are significantly more expensive than input tokens, and developers often hesitate to let models return long verbatim quotes. Making citations financially free incentivizes developers to treat Claude as a primary verification engine, anchoring enterprise workflows firmly inside Anthropic's ecosystem.