Gemini API adds Office files and richer tool responses
Google expanded Gemini API file handling on November 5, 2026, adding PPT, ODP and ODS inputs and visual processing for older DOC files. Tools can also return Office documents, audio and video to the model, including references to uploaded files rather than inline bytes.
For developers, the useful question is what information survives ingestion. File acceptance alone is a poor acceptance test.
What Gemini actually reads from each format
Google’s document-processing guide describes four distinct paths:
| File format | What reaches the model |
|---|---|
| DOC, PPT, ODP | Every page or slide is rendered as an image, including visual layout. |
| DOCX, DOTX, ODT, PPTX | Whole-file text. DOCX and PPTX also include embedded images, but pages are not rendered. |
| XLSX, XLS | First sheet only, converted to CSV values without formulas or formatting. |
| ODS | First sheet only, including cell data and formulas. |
Other workbook sheets are ignored. Move the needed sheet first or save it separately. The same processing applies to direct inputs and tool results. Existing file and page limits apply to the whole file. The guide specifies 50 MB or 1,000 pages for PDFs.
For a finance workflow, test a known total from a non-first sheet. For a presentation, ask a question that depends on chart placement. These checks can expose a mismatch between the task and its input.
Tools can hand back the source material
The release notes distinguish the two API shapes. GenerateContent function response parts now accept Office and OpenDocument files, audio and video. The new fileData support can carry a Files API URI. Both inlineData and fileData accept displayName, while video parts accept mediaProcessing with STATIC or AGENTIC values.
For the Interactions API, the function-calling guide documents document, audio and video blocks inside a function_result step’s result. A media block carries base64 data or a uri. Video blocks are available in v1beta and v1alpha. Their processing value can be static or agentic. These field names and capitalization are API-specific, so a payload should not be copied unchanged between the two interfaces.
The guide scopes multimodal function responses to Gemini 3 series models and demonstrates gemini-3.8-flash. Unsupported MIME types produce a 400 INVALID_ARGUMENT error. This is a documented compatibility path, not a promise that every model and API version accepts every new combination.
For example, a retrieval tool could return the actual deck and a related video for the next model turn instead of reducing both to a text summary. That is a useful design option when a summary would discard evidence. It still leaves the application responsible for choosing the right file, enforcing access permissions and checking the answer.
File URIs do not remove size and lifetime limits
Google recommends a Files API URI for large tool results because inline data counts toward request size. Its Files API documentation lists 20 GB of storage per project, a 2 GB per-file ceiling and automatic deletion after 48 hours. Uploaded-file metadata can be retrieved, but the files cannot be downloaded through that API. These storage ceilings do not override the tighter processing limits for a document type.
Keep your own authorized original and plan for an expired reference in a long-running job. A retry should be able to locate the correct source again, rather than silently substitute a different version. ByteForward’s Gemini API CLI guide covers the separate shell-client route for scripted workflows, which has its own client limits.
Budget for the chosen model and validate a small sample
The Gemini 3.8 Flash model page lists a 1,048,576-token input limit and 65,536-token output limit. Those are model ceilings, not a guarantee that a particular document will be understood completely.
At the current Standard paid rate, Gemini 3.8 Flash costs $0.75 per million input tokens and $3.75 per million output tokens, including thinking, through December 31, 2026. The listed rates rise to $1.50 and $7.50 on January 1, 2027. Other models and service tiers have their own rates.
Before adopting the new path, use a small, non-sensitive fixture with known answers. Include a second worksheet, a formula and a visually meaningful slide so omissions are easy to detect. Compare what the application sends with what the answer cites, record token usage, and test an expired file reference. These are suggested acceptance checks, not ByteForward benchmark results. We reviewed the release notes and documentation but did not run the API.
Featured image: AI-generated ByteForward editorial illustration with the official Gemini wordmark. Gemini is a trademark of Google LLC. ByteForward is not endorsed by or affiliated with Google. The document objects are conceptual, not a product interface.
