Gemini file upload limits in 2026: size, count and file types

These are the Gemini app limits from Google's Gemini Apps Help, checked in October 2026. Google notes that the 10-file limit is subject to availability.

You can attach up to 10 files to one prompt, and when you upload a folder, each file inside counts toward those 10. Two bundle types have their own caps:

• Code folder or GitHub repository: one per chat, up to 5,000 files and 100 MB in total.
• ZIP file: up to 10 files and 100 MB, with no video or audio inside.

Google publishes no numeric daily file cap. Instead, a rolling limit on chats with files resets after some time, and Google AI Pro and Ultra raise it.

Passing the 100 MB check does not mean Gemini reads your whole file. The content has to fit in your plan's context window, and Google warns that going over it can lead to answers that miss details:

• Without a Google AI plan: 32K tokens
• Google AI Plus: 128K tokens
• Google AI Pro and Ultra: 1 million tokens, which Google puts at about 1,500 pages of text or 30,000 lines of code
• Deep Think: 192K tokens

For CSV and JSON, the token window is almost always the real bottleneck. A file far below 100 MB can exceed 32K or 128K tokens, and JSON costs more than CSV because every key name repeats on every record.

rows.page runs DuckDB-WASM in your browser tab, so opening a file needs no account and nothing is uploaded unless you press Share.

1. Drop your file on rows.page, or click Paste data to paste cells copied from Excel or Google Sheets. It opens .csv, .txt, .tsv, .tab, .json, .jsonl, .ndjson, .ldjson, .parquet and .pq files.
2. Check the Columns panel for each column's type, empty values, min and max, and a histogram or top 10 values. Hide columns your prompt doesn't need.
3. Narrow rows with the filter box, for example country:NL, price:>20 or -name:* (empty), or click a histogram bar.
4. Export the view as CSV, JSON, JSONL, Parquet or Markdown (first 1,000 rows) and attach that file to Gemini.

A tab holds up to 1.5 GB of CSV, TSV, JSON or JSONL. Bigger CSV, TSV and JSONL files load partially, a bigger .json file is refused, and Parquet skips that check. rows.page does not read Excel files, so save them as CSV or paste the cells.

Instead of sending raw rows, aggregate the dataset and upload only the summary. Type SQL that starts with SELECT, WITH or FROM into the filter box; the table is called data.

SELECT
  "user.address.city" AS city,
  count(*) AS records,
  avg(price) AS avg_price,
  median(price) AS median_price
FROM data
WHERE price IS NOT NULL
GROUP BY 1
ORDER BY records DESC

• Nested JSON flattens into dot-path columns, up to 8 levels deep and 200 child fields per object. Quote names that contain dots, such as "user.address.city".
• FROM (SUMMARIZE data) returns a statistical profile of every column, a compact summary to hand to Gemini.
• Remove duplicates with SELECT DISTINCT * EXCLUDE (__rid) FROM data, since the hidden __rid row number makes every row unique.
• SQL results export up to 1,000,000 rows, while filtered views export every matching row.

To send a dataset to a teammate or another tool, press Share or push it from your terminal:

curl -T data.csv https://rows.page

Without an account, a push can be up to 100 MB, you get 50 pushes and 1 GB per IP address per UTC day, and links expire after 7 days. Each hosted dataset also gets a raw URL at https://rows.page/<id>/raw.<ext>, and MCP clients such as Claude Code or Cursor can read it through https://rows.page/mcp.