What changed on 18 November

Google's Gemini API changelog for 18 November 2026 lists two additions to file handling. Word, PowerPoint and Excel files, OpenDocument formats included, now count as model input. A request can also name exactly which part of a file the model sees: a page range, a worksheet, a block of cells, a time interval in audio or video, or a content layer. The same entry adds the slicing options to function responses, so a tool your agent calls can return part of a file instead of the whole thing.

The update builds on the 5 November release, which added PPT, ODP and ODS input and let tools return Office files, audio and video. For anyone building document agents over board packs, tender documents or council spreadsheets, the November changes mean a 300-slide deck no longer has to be converted to PDF and sent whole on every turn.

How the slicing works

Office files go into a request as a document content block carrying the file's MIME type. A Files API URI or inline data both work. When you add a processing object with "type": "static", the document processing guide says Gemini counts and bills only the slice you asked for. The model is also told the file's total page count and worksheet names, which lets it ask for a different range on a later turn.

Pages in PDFs, Word files and slide decks are selected with page_options. The start_page and end_page fields are 1-indexed and inclusive. The content_layers field takes "page_image", "page_text" or both, and it defaults to images once processing is set. Spreadsheets use spreadsheet_options instead. That object takes sheet_index (the tab's position, starting at 1), visible_cells in A1 notation such as "A1:H50", and layers called "data" (evaluated values), "formulas" (the raw formulas) and "visual" (a rendered snapshot of charts and layout). The spreadsheet default is data plus formulas. Resolution ("low", "medium" or "high") sits on the outer document block, outside the options objects.

The defaults also differ between model generations. The guide's comparison table shows Gemini 4 models rendering .docx and .pptx pages as images at about 560 tokens a page, while Gemini 3 and 2.5 extract the text and embedded images. For .xlsx files, Gemini 4 reads values, formulas and sheet names, and the older generations get every sheet as CSV values with no formulas. Code that moves from Gemini 3 to Gemini 4 will therefore see different token counts for the same file.

Building a spreadsheet audit step

The steps below follow the official examples, which use the google-genai Python SDK, the Interactions API and gemini-4-hydrogen.

  1. Upload the workbook with client.files.upload. According to the guide, uploads are stored for 48 hours.
  2. Send a document block with spreadsheet_options that name the tab and the cell range.
  3. Read interaction.output_text.

The documentation's spreadsheet example, trimmed slightly:

workbook = client.files.upload(file="financial_model.xlsx")

interaction = client.interactions.create(
    model="gemini-4-hydrogen",
    input=[
        {
            "type": "document",
            "uri": workbook.uri,
            "mime_type": "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet",
            "resolution": "high",
            "processing": {
                "type": "static",
                "spreadsheet_options": {
                    "sheet_index": 2,
                    "visible_cells": "A1:H50",
                    "content_layers": ["data", "formulas", "visual"],
                },
            },
        },
        {"type": "text", "text": "Audit the EBITDA formulas and chart in Sheet 2 (A1:H50)."},
    ],
)
print(interaction.output_text)

For long decks and reports, the guide suggests a two-pass pattern. The first pass reads everything with content_layers: ["page_text"]. The second pass requests "page_image" at "high" resolution only for the slides or pages where charts or dense layouts matter.

Letting the agent choose the pages

The more useful pattern for agents is in the function calling guide. Its example defines an inspect_document_pages tool with integer start_page and end_page parameters. When the model calls the tool, your code passes the model's arguments straight into page_options and returns the uploaded file inside a function_result:

"processing": {
    "type": "static",
    "page_options": {
        "start_page": call.arguments["start_page"],
        "end_page": call.arguments["end_page"],
        "content_layers": ["page_image"],
    },
},

The second turn links to the first with previous_interaction_id. Because the model already knows the page count, it can read a contents page and then pull the sections it needs, which suits retrieval over long manuals where chunking and embedding the whole file would be overkill. The guide states that processing options work on Gemini 3 and later models, both in user input and inside function_result.result.

Limits and costs

Several limits apply to the full uploaded file, regardless of how small the slice is. The guide sets 50 MB per file, 1,000 pages per PDF, 2,000 pages or slides per Office or OpenDocument file, and 50,000 cells per sheet. A start_page of zero or less, or a sheet_index past the last tab, returns 400 INVALID_ARGUMENT. A start_page beyond the end of the document returns an empty slice without any error, so an agent can mistake a bad range for a page with nothing on it.

Page images cost about 280 tokens per page at low resolution, 560 at medium and 1,120 at high on Gemini 3 and 4. The pricing page lists Gemini 4 Hydrogen input at US$1.50 per million tokens on the paid tier, with free input on the free tier. By our arithmetic from those figures, a 40-slide deck at medium resolution comes to about 22,400 tokens, or roughly US$0.034 of input per request. Sending five slides at high resolution costs about 5,600 tokens. Gemini 4 Neon input costs US$3.00 per million, and neither Neon nor Carbon has a free tier.

Rohit Ramachandran's write-up from 18 November makes a point NZ teams handling personal information should take seriously: "selecting less for the model is not the same as uploading less to Google." Slicing happens on Google's side after the whole file has been uploaded. Content that must stay within your systems under the Privacy Act needs removing before upload. The same post warns that sheet positions can change between versions of a workbook, so sheet_index: 2 may point at a different tab next month. A returned formula also tells you nothing about whether the value was recalculated.

The documentation has a few rough edges. The REST samples switch between /v1/interactions and /v1beta/interactions, and the guide doesn't say whether Gemini 2.5 accepts processing options at all. Gemini 2.5 Pro and 2.5 Flash are due to be routed to newer models on 1 December, so new builds should target Gemini 3 or later. If you build a document agent on slicing, you can share it on the ASAI projects page.

Sources

Build with people working on the same things

Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.

Become a member