Claude Accuracy: How to Reduce Errors in Long Documents

TL;DR

A 1M-token window is not a 1M-token working range. Anthropic names the degradation context rot, publishes no error rate, and documents six controls that reduce errors over long files.

A million-token context window is not a million-token working range. Anthropic's own documentation says that "as token count grows, accuracy and recall degrade" and gives the effect a name: context rot. It publishes no error rate for long-document work, and it does publish six controls that reduce the error — put the document above the question, structure it with tags, extract quotes before answering, allow "I don't know", turn on sentence-level citations, and restrict the answer to the supplied text.

Everything below is drawn from Anthropic's technical documentation at platform.claude.com and three peer-reviewed or preprint papers, all retrieved 2026-09-18. Model facts move quickly: context windows, cutoffs and feature availability are as documented on that date, so pin your reading to it and re-check before you rely on a number.

Key Takeaways

Anthropic names the failure itself. Its context-window page states that "as token count grows, accuracy and recall degrade, a phenomenon known as context rot" (Anthropic, Context windows, retrieved 2026-09-18, SOURCED).

Position matters more than length. Model performance "significantly degrades when models must access relevant information in the middle of long contexts, even for explicitly long-context models" (Liu et al., TACL, arXiv:2307.03172, retrieved 2026-09-18, SOURCED).

Put the document first and the question last. Anthropic's prompting reference says queries at the end "can improve response quality by up to 30 percent in tests, especially with complex, multidocument inputs" (Anthropic, Prompting best practices, retrieved 2026-09-18, SOURCED).

Quote first, then answer. For documents over 20k tokens, Anthropic's guidance is to "extract word-for-word quotes first before performing its task" (Anthropic, Reduce hallucinations, retrieved 2026-09-18, SOURCED).

No control closes the gap. The same page states these techniques "don't eliminate them entirely" and tells you to "always validate critical information, especially for high-stakes decisions" (retrieved 2026-09-18, SOURCED).

What Anthropic documents, and what it does not publish

Anthropic publishes no accuracy percentage, hallucination rate or citation-error rate for answering questions from long documents. We looked through its models overview, context-window documentation, citations reference and hallucination guidance and found none, so none is given here.

What it does publish is a mechanism and a warning. The context-window page describes the window as the model's working memory, distinct from the corpus it was trained on, and then says plainly that "more context isn't automatically better. As token count grows, accuracy and recall degrade, a phenomenon known as context rot. This makes curating what's in context just as important as how much space is available" (Anthropic, Context windows, retrieved 2026-09-18, SOURCED). The hallucination guidance is equally direct about the model class: "Even the most advanced language models, like Claude, can sometimes generate text that is factually incorrect or inconsistent with the given context" (Anthropic, Reduce hallucinations, retrieved 2026-09-18, SOURCED).

One capacity figure sets the scale of the problem. Current Claude models carry a 1M-token context window, Claude Sonnet 4.5 and Claude Haiku 4.5 carry 200K, and 1M tokens is "roughly 555k words" on the current tokenizer (Anthropic, Models overview, retrieved 2026-09-18, SOURCED). A 555,000-word capacity invites you to paste a filing cabinet in. The context-rot warning sits in the same documentation, telling you not to.

Four reasons a long document produces a wrong answer

1. Where the fact sits changes whether it is used

Liu et al. tested multi-document question answering and key-value retrieval and found that "performance is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts, even for explicitly long-context models" (Liu et al., TACL, arXiv:2307.03172, retrieved 2026-09-18, SOURCED).

Levy et al. isolated length itself by padding the same questions to different lengths, and report "a notable degradation in LLMs' reasoning performance at much shorter input lengths than their technical maximum" (Levy, Jacoby and Goldberg, arXiv:2402.14848, retrieved 2026-09-18, SOURCED).

Neither study evaluated the current Claude models, so read them as the shape of the effect rather than a measurement of today's behaviour (ANALYSIS). The shape is what matters: a clause on page 200 of a 300-page contract sits in the worst position in the file.

2. Retrieval decides what the model ever sees

When a document is served through a retrieval layer rather than pasted whole, the answer can only be as good as the chunks that came back. Barnett et al. report seven failure points from three production retrieval systems and conclude that "validation of a RAG system is only feasible during operation" (Barnett et al., arXiv:2401.05856, retrieved 2026-09-18, SOURCED). A pipeline that tests well on sample questions can still miss the clause your reviewer asks about next week, and nothing in the answer will say so.

3. A summary of a summary loses the source

Long conversations about long files eventually get compressed. Anthropic's compaction documentation states the reason directly: "as a conversation grows, response quality degrades, so compaction replaces older content with a concise summary". It describes the mechanics too: the API "automatically drops all content blocks prior to the compaction block, continuing the conversation from the summary" (Anthropic, Compaction, retrieved 2026-09-18, SOURCED).

That is a sensible trade for a long task, and it changes what a later answer rests on. Once the document blocks are dropped, the model is answering from its own summary rather than the file, and anything the summary flattened is gone without the answer mentioning it (ANALYSIS).

4. The file contradicts itself

One PDF can hold a 2024 price list, a 2026 amendment and a superseded schedule that was never removed. Nothing in a retrieval or attention mechanism adjudicates between them, and a confident answer can quote the wrong one accurately. This is the failure our own corrections work turns up most often in practice, and it is not documented by any provider because it is a property of your files, not of the model (ANALYSIS). The method for taking a claim back to its source is in how to trace an AI claim back to its original source.

Six controls Anthropic documents

Control 1 — Put the document at the top, the question at the bottom

Anthropic's prompting reference, in its long-context section for inputs of "20k+ tokens", says: "Put longform data at the top: Place your long documents and inputs near the top of your prompt, above your query, instructions, and examples. This improves performance across all models." It adds that queries at the end "can improve response quality by up to 30 percent in tests, especially with complex, multidocument inputs" (Anthropic, Prompting best practices, retrieved 2026-09-18, SOURCED).

This is the cheapest change on the list and the one most prompts get backwards.

Control 2 — Wrap each document with its source metadata

The same section tells you to "wrap each document in <document> tags with <document_content> and <source> (and other metadata) subtags for clarity", nesting them inside <documents> with an index per document (retrieved 2026-09-18, SOURCED). The gain is attribution: when each document carries a filename and an index, an answer that mixes two files is visible as soon as you read which source it names (ANALYSIS).

Control 3 — Make it quote before it answers

Anthropic's hallucination guidance: "For tasks involving long documents (>20k tokens), ask Claude to extract word-for-word quotes first before performing its task. This grounds its responses in the actual text, reducing hallucinations." Its prompting reference gives the same instruction, explaining that quoting first "helps Claude focus on the relevant content and ignore the rest of the document" (both retrieved 2026-09-18, SOURCED).

Anthropic's own example asks for quotes inside <quotes> tags first, then the analysis built only on those quotes. Run in that order, a fabricated quote is visible before it reaches your answer.

Control 4 — Give it permission to fail

Anthropic's first listed strategy is to allow Claude to say "I don't know" — "explicitly give Claude permission to admit uncertainty. This simple technique can drastically reduce false information" (Anthropic, Reduce hallucinations, retrieved 2026-09-18, SOURCED). The same page pairs this with external knowledge restriction: "explicitly instruct Claude to only use information from provided documents and not its general knowledge".

For a document task both belong in the prompt. Without them, a question about a clause that is not in the file gets answered from training data, in the same tone as an answer that is in the file.

Control 5 — Turn on citations and read the span

The Citations feature returns "the exact passages that support each claim, so you can verify answers and surface sources to your users." Documents are chunked so that "sentence chunking lets Claude cite a single sentence or chain together multiple consecutive sentences to cite a paragraph or longer passage." Citations arrive as page ranges for PDFs, character index ranges for plain text, and content-block indices for custom content (Anthropic, Citations, retrieved 2026-09-18, SOURCED).

Two documented limits belong in the same breath: "Only text citations are currently supported. Image citations are not yet possible." And "PDFs that are scans of documents and do not contain extractable text are not citable" (same page, retrieved 2026-09-18, SOURCED). A scanned contract is exactly the document you most want a citation from.

For answers assembled from your own retrieval layer, search result content blocks "let Claude cite your own content the same way it cites web search results", with the source and title you supplied (Anthropic, Search results, retrieved 2026-09-18, SOURCED).

Control 6 — Verify each claim, and retract what has no quote

The last control is a second pass rather than a setting. Anthropic's guidance: "Make Claude's response auditable by having it cite quotes and sources for each of its claims. You can also have Claude verify each claim by finding a supporting quote after it generates a response. If it can't find a quote, it must retract the claim" (retrieved 2026-09-18, SOURCED).

Its worked example is blunter than most house styles would dare: remove the unsupported claim and leave empty brackets where it was. That is the same discipline we hold our own drafts to in how to stop ChatGPT from making things up, applied inside a single document task.

What none of this fixes

Anthropic states the limit of its own advice: these techniques "significantly reduce hallucinations" but "don't eliminate them entirely. Always validate critical information, especially for high-stakes decisions" (Anthropic, Reduce hallucinations, retrieved 2026-09-18, SOURCED).

Three other limits are documented rather than fixable:

Scans. A PDF with no extractable text cannot be cited, so your verification step has nothing to open. Convert it with optical character recognition first, and check the conversion.
Size before the page limit. A request can carry up to 600 PDF pages, or 100 when the request's context window is under 1M tokens, and "dense PDFs (many small-font pages, complex tables, or heavy graphics) can fill the context window before reaching the page limit" (Anthropic, PDF support, retrieved 2026-09-18, SOURCED).
Knowledge cutoffs. Reliable knowledge cutoffs for current models run from Feb 2025 for Claude Haiku 4.5 to Jun 2026 for Claude Fable 5.1 (Anthropic, Models overview, retrieved 2026-09-18, SOURCED). Anything after a model's cutoff must come from the document you supplied, which is one more reason to restrict the answer to it.

The general version of this problem — a correct source summarised wrongly — is covered in what "99% accurate" actually means, and the fast check on any single answer is in how to check if ChatGPT is right.

The order we would work in

Cut the input before you improve the prompt. Send the three relevant sections, not the 300-page file. Curating context is Anthropic's own advice on the context-rot page (ANALYSIS).
Document first, question last. One reordering, documented gain.
Quote, then answer. Ask for the quotes in tags, read them, then accept the analysis.
Citations on, and open one. Check a span from late in the document, where attention is thinnest.
Re-run the same question on the same file. Anthropic lists best-of-N comparison among its advanced techniques; inconsistency between runs "could indicate hallucinations" (retrieved 2026-09-18, SOURCED).

How This Guide Was Sourced

Written and maintained by the LogicBalls editorial team (logicballs.com). Disclosure: LogicBalls builds AI writing tools and routes work to models from several providers, including Claude-class models. This post describes documented behaviour and controls for long-document work. It is not a purchase comparison, it names no alternative and it recommends none.

Sources. Anthropic technical documentation at platform.claude.com: context windows, models overview, prompting best practices (long-context section), reduce hallucinations, citations, search results, compaction and PDF support. Research: Liu et al. (arXiv:2307.03172, TACL), Levy, Jacoby and Goldberg (arXiv:2402.14848) and Barnett et al. (arXiv:2401.05856). All retrieved 2026-09-18, linked inline, and every quotation was matched word for word against the page text on that date.

Pin your reading. Context windows, cutoffs, page limits and feature availability change on a scale of weeks. Every figure above carries its retrieval date; treat any of them older than a month as unverified.

What could not be verified. Anthropic publishes no accuracy, hallucination or citation-error rate for long-document tasks, so this post gives none. The "up to 30 percent" figure is Anthropic's own reported test result and we could not inspect the underlying tests. None of the three papers evaluated the current Claude models named here. We ran no test of our own and publish no measurement of how much any control helps.

AI involvement. AI assisted with research and drafting. Every quotation and URL was checked against the live source page on 2026-09-18 before publication, and the post was reviewed against this blog's pre-publication gate before it was scheduled.

No LogicBalls telemetry is used in this guide. Every figure above is external and linked.

Frequently Asked Questions

How accurate is Claude with long documents?

No published figure answers that. Anthropic reports no error rate for long-document tasks and states that accuracy and recall degrade as the token count grows. What it offers instead is a set of controls and an instruction to validate critical information yourself.

Does a 1M-token context window mean I can paste a whole book?

You can, and Anthropic's own documentation advises against treating capacity as a working range: "more context isn't automatically better." Research on input length finds reasoning degrading well before a model's technical maximum.

Do citations prove the answer is right?

They prove a passage exists and show you where it is. The check is opening the cited span and reading it. Citations also do not cover images, and a scanned PDF with no extractable text cannot be cited at all.

Where in a long file do errors cluster?

The middle is the weakest position in the published research, with performance highest at the beginning and end. When you spot-check an answer, check a fact from the middle and one from late in the document.

Is it better to paste the document or to retrieve chunks?

Pasting the relevant sections keeps the source in front of the model. Retrieval scales further and adds its own failure points, which Barnett et al. conclude only become visible in operation. Either way, the verification step is the same: open the cited passage.

Conclusion

Long-document accuracy is a curation problem before it is a prompting problem. Anthropic documents the degradation, names it, and publishes six controls that work against it — none of which it claims will eliminate the error. Send less, put the document above the question, make the model quote before it answers, let it say it does not know, turn citations on, and open one span before you act on the answer.