Long Context Reasoning

Long Context Reasoning

Quick answer: Long Context Reasoning evaluates an AI model's ability to perform multi-step reasoning across very long input sequences — testing whether models can retain relevant information, make multi-hop inferences, and synthesise answers from content spread across large documents or conversation histories. Fugu scores 74.7% and Fugu Ultra scores 73.3% as of June 2026.

At a Glance

What it tests: Reasoning quality over long contexts — including retrieval of relevant facts from distant positions, multi-hop inference chains, and synthesis of information across multiple sections of a long document.

Why it matters: Many enterprise AI use cases involve long documents, codebases, or conversation histories. A model that reasons accurately over long contexts is far more useful in production than one that degrades with length.

Known limitations: Performance is sensitive to context window size and attention architecture. Results may vary across different long-context task types (e.g., book QA vs. code reasoning vs. multi-document summarisation).

What Long Context Reasoning Measures

Long Context Reasoning benchmarks test a model's ability to maintain accurate reasoning over inputs that span tens of thousands to hundreds of thousands of tokens. Tasks typically require the model to retrieve specific facts from early in a long document, connect them with information elsewhere, and reason to a correct conclusion — rather than simply finding and repeating a nearby passage.

A score in the 70–75% range represents frontier performance as of mid-2026 and indicates strong long-context capability. Models that perform well here can be trusted to reason accurately over full codebases, legal documents, or extended research papers within their context window.

Benchmark Specifications

State-of-the-Art Results

Scores sourced from Sakana AI's Fugu technical report, June 2026.