All capabilities Data and analysis

Fit a whole codebase in one prompt, and find its limit

A whole codebase, a full deposition, an hour of video. The window is real; the retrieval behaviour inside it is the part people get wrong.

DocsDocumented in partTestNot tested by benchr Quick · API Docs checked September 1, 2026

Gemini APIClaude APIGemini 3.1 ProGemini 3.7 Flash+2

Overview

Accepts context up to 1M tokens on the documented Gemini models, against 8k and 32k windows in the recent past.
Covers text, video and audio use cases explicitly - summarisation, question answering, agentic workflows, content moderation, meeting summaries.
Removes the retrieval layer for problems that fit, which removes a whole class of chunking bugs.

01Ask one question per request. Documented accuracy is highest for a single retrieval and degrades when a request needs multiple separate facts.
02Put the corpus first and the question last, and keep the corpus byte-identical across requests if you also want prompt caching to hit.
03For multi-fact work, run several single-needle passes and join the answers yourself rather than asking for all of them at once.
04Measure latency, not just cost. Longer queries increase latency, which the documentation states plainly.

Multiple needles in one query reduce accuracy compared with a single retrieval - this is documented, not folklore.
Performance varies by context and longer queries increase latency.
The documentation frames long context as a trade-off between retrieval accuracy and cost efficiency, not a free upgrade.
A 1M window on the spec sheet is not the same as usable recall across the whole window.

benchr has not run this capability. No test has been published.

benchr test runs →

Source-check history

September 1, 2026DocsDocumented in partFirst logged as Partial. The window sizes and use cases are documented; the accuracy behaviour is described qualitatively with no published curve, so the record does not claim one.