---
source: https://iinventedlight.com/library/notes/ao-context-million-token-windows.md
title: "Million-token context windows: what Gemini 1.5 reported"
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
source_license: "CC BY 4.0; restated in our own words"
attribution: "Summary by iinventedlight (AI-assisted). Sources: arXiv, [2403.05530] Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context, CC BY 4.0; arXiv, [2401.18059] RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval, CC BY 4.0."
author: "iinventedlight (synthesis of cited sources); text drafted with AI assistance"
date: 2026-11-04
status: published
checked_on: 2026-11-04
as_of: 2026-11-04
confidence: medium
untested: "No experiment was run; only the fetched arXiv abstract pages were read, so results are as reported by the authors."
category: agent-operations
topic_area: "agent-ops: context and retrieval"
tags: ['agent-operations', 'context', 'long-context']
sources:
  - id: S1
    title: "[2403.05530] Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context"
    publisher: arXiv
    url: https://arxiv.org/abs/2403.05530
    accessed: 2026-11-04
    license: "CC BY 4.0"
  - id: S2
    title: "[2401.18059] RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval"
    publisher: arXiv
    url: https://arxiv.org/abs/2401.18059
    accessed: 2026-11-04
    license: "CC BY 4.0"
---
# Million-token context windows: what Gemini 1.5 reported

**Answer:** Google's Gemini 1.5 report claims near-perfect (over 99%) retrieval up to at least 10 million tokens, far beyond the 128k to 200k windows of models it compared against. Long windows compete with retrieval methods such as RAPTOR rather than fully replacing them.

## What the report says

Gemini 1.5 Pro and the lighter Gemini 1.5 Flash are described as able to recall and reason over millions of tokens of text, video and audio. The authors report near-perfect recall on long-context retrieval tasks across modalities and continued improvement in next-token prediction as context grows.

## Long context vs retrieval

The RAPTOR paper (S2) takes the other route: keep the prompt small and retrieve from a tree of summaries. Its authors argue flat chunk retrieval limits holistic understanding, which is the same problem long windows try to solve by brute force.

## Operator trade-offs

- Long prompts cost more per call and add latency.
- Needle-style retrieval tests do not prove the model reasons well over everything.
- Retrieval keeps cost down and can cite exact passages.

## Claims
- id: C1
  claim: Gemini 1.5 models recall and reason over millions of tokens of context.
  source: S1
  evidence: "capable of recalling and reasoning over fine-grained information from millions of tokens of context"
- id: C2
  claim: The report finds near-perfect retrieval above 99% up to at least 10M tokens.
  source: S1
  evidence: "near-perfect retrieval (>99%) up to at least 10M tokens"
- id: C3
  claim: Gemini 1.5 Flash is a lightweight variant designed for efficiency.
  source: S1
  evidence: "Gemini 1.5 Flash, a more lightweight variant designed for efficiency"
- id: C4
  claim: RAPTOR retrieves from a tree of summaries at inference time.
  source: S2
  evidence: "At inference time, our RAPTOR model retrieves from this tree"

## Sources
- [S1] [2403.05530] Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv. https://arxiv.org/abs/2403.05530 . License: CC BY 4.0. Accessed 2026-11-04.
- [S2] [2401.18059] RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval. arXiv. https://arxiv.org/abs/2401.18059 . License: CC BY 4.0. Accessed 2026-11-04.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted).