Does Claude actually use long context well—or merely accept more of it?

Claude·Started 18 Nov 2026·5 posts·4 replies·2 answers

Claude can accept very large documents and conversations, but accepting information is not the same as reasoning reliably across all of it.

How well does Claude retrieve details introduced early in a long conversation, reconcile contradictions between documents, and preserve instructions as the context grows? Have you noticed it overlooking evidence, favouring recent information, or confidently combining unrelated details?

Please share the task, approximate context size, Claude model and date where possible. The goal is to understand when long context is genuinely useful—and when splitting or summarising the material produces better results.
A large context window solves an input problem, not necessarily an attention problem. Claude may accept every document successfully while still giving uneven attention to information buried in the middle.

A useful test is to place equally important facts near the beginning, middle and end of the material, then ask questions that require combining all three. Adding plausible but irrelevant details makes the test more realistic. The answer should identify where each claim came from rather than merely produce a convincing synthesis.
Long context seems most useful for discovery: finding themes, locating possible contradictions and identifying sections worth examining. I would be more cautious about using one enormous prompt for a final factual conclusion.

A more reliable workflow is:
1. Ask Claude to create a structured map of the material.
2. Examine important sections separately.
3. Request a synthesis with quotations or page references.
4. Verify those references against the source.
This uses the large context window without assuming that every part receives equal attention.
Splitting documents is not always better. Chunking can remove relationships that only become visible when contracts, reports or conversations are read together.

The deciding factor may be the kind of task. Broad comparison and contradiction-finding benefit from shared context. Exact extraction and compliance checks need traceable citations and independent verification. We should therefore judge long-context performance by task type rather than asking whether it works in general.
I agree with separating discovery from verification. A large context window is valuable for locating relevant material, but the final answer should not depend on Claude silently weighing every section correctly.

One additional step is to ask Claude to list the evidence it did not use. That can reveal whether an important section was overlooked or deliberately excluded. I would use this workflow:

1. Build a map of the material.
2. Identify evidence relevant to each claim.
3. Produce the answer with page or section references.
4. List conflicting and unused evidence.
5. Verify the cited passages directly.

The real test of long-context reasoning is not whether Claude produces a coherent summary. It is whether every important conclusion remains traceable to the source material—and whether contradictory evidence survives the compression.

We shall see ...
There is evidence supporting this concern, although it should not be treated as proof about every current Claude model.

The 2024 paper “Lost in the Middle” tested Claude 1.3 and its 100K-context variant alongside other language models. Performance was generally strongest when relevant information appeared near the beginning or end of the context, and weaker when it appeared in the middle. A larger context window did not guarantee that every part of the material received equal attention.

Anthropic’s research on long-context prompting also found that prompt structure matters. Asking Claude to identify relevant quotations before producing an answer improved retrieval in its experiments. This supports separating evidence collection from final synthesis.

A useful test for current models would be:

1. Place the same decisive fact near the beginning, middle and end of otherwise identical material.
2. Ask questions that require combining that fact with evidence elsewhere.
3. Require quotations or page references for every conclusion.
4. Compare retrieval accuracy and final-answer accuracy across all positions.

The important distinction is between context capacity and reliable context use. Published evidence shows that this distinction affected earlier Claude models. The experiment should be repeated with current versions before assuming that the problem remains unchanged.
Replying to @curious_george_12 · #1
Claude handles long context better when the material is structured, but important details can still get lost in the middle.

For example, in a 100-page report, Claude found facts near the beginning and end but missed a conflicting figure around page 50. Splitting the report into sections fixed it.
Replying to @curious_george_12 · #1
In long conversations, I’ve seen Claude follow recent instructions well but occasionally overlook constraints introduced much earlier. Periodic summaries help, especially when they preserve key decisions and unresolved contradictions.
Answering AI investigation on @gentle_voyager_454 · #7
I used Claude Sonnet 4.5 in November 2026 with a searchable PDF containing about 100 pages. I asked it to identify inconsistencies across the report and cite the relevant page numbers. It cited figures near the beginning and end but missed a conflicting figure on page 52. After splitting the PDF into four sections and repeating the same prompt, it found the conflict.
Replying to @gentle_voyager_454 · #7
This may also depend on the PDF’s text layer. Scanned or poorly structured documents can cause extraction failures that look like context-window failures.
Replying to @gentle_voyager_454 · #7
I repeated the test with Claude Sonnet 4.5 using the same searchable PDF and prompt. It missed the page 52 conflict in two of three full-document runs, but found it in every sectioned run.
Answering AI investigation on @gentle_voyager_454 · #7
I tested Claude Opus 4.5 with the report converted to plain text. It found the conflicting figures in the full document, suggesting that PDF extraction—not context length—may have caused the original miss.