Google has expanded the Gemini API’s document-processing capabilities with native support for Microsoft Office and OpenDocument files, alongside controls for selecting only the portions of a file that a model should process.
Applications can submit Word documents, PowerPoint presentations, Excel workbooks and OpenDocument files without first converting them into another supported format. They can also target document pages, worksheet tabs, cell ranges, audio or video intervals, and specific content layers. Google says customers process and pay only for the selected content.
The combination matters more than basic file-format support. It turns file selection into part of the inference request and its billing semantics, potentially reducing the amount of irrelevant material sent to a model. It also creates a new compatibility challenge for gateways that try to present a consistent interface across providers.
File conversion is no longer always required
Document applications commonly use conversion services to extract text, render pages or normalize files before sending their contents to a model. Native Office support can remove some of those steps from Gemini-based pipelines, particularly for workflows built around reports, presentations and spreadsheets.
That could simplify ingestion systems and preserve more of a document’s original structure. A financial assistant, for example, could direct Gemini to a particular worksheet or cell range instead of extracting and transmitting an entire workbook. A presentation workflow could select a group of slides, while a media application could isolate a relevant audio or video interval.
Slicing is available both in user prompts and in file-bearing function responses. The latter is significant for agents: a tool can return a file while identifying the portion that should be considered by the model, rather than forcing the application to create a separate derivative file for every tool result.
The practical gains will vary. Complex formulas, embedded objects and unusual Office layouts may not behave consistently across files, and the research package does not establish universal token or latency reductions. Teams will need representative tests before removing existing conversion and validation stages.
Selected content becomes a billing boundary
Google’s statement that customers pay only for selected content introduces a more granular cost boundary than whole-file processing. Applications can potentially avoid paying to inspect hundreds of irrelevant pages or unused spreadsheet tabs when a task concerns only a narrow section.
Developers should still avoid treating a selector as a guaranteed cost estimate. The eventual token count and processing time can depend on file structure, the selected slice and how the model represents that material. Usage records should retain both the original file reference and the exact selection parameters so that costs can be explained later.
This is especially relevant to resellers and software platforms that show customers a usage breakdown. Their ledgers may need to distinguish the size of an uploaded file from the portion submitted for inference. Without that distinction, an invoice could appear inconsistent with the user-visible document or make cost disputes difficult to investigate.
OpenAI compatibility does not cover every selector
Page, worksheet, cell-range, media-interval and content-layer selectors do not map neatly onto a conventional OpenAI-compatible file input. A multi-model API gateway for teams therefore has several choices: expose Gemini-specific extensions, translate a portable subset of selectors, or require applications to preprocess files before using providers that lack equivalent controls.
Silently discarding unsupported selectors would be risky. It could cause a provider to process an entire document, changing both the answer and the bill. Gateways should validate requests, identify unsupported selection types and record the normalized request that was ultimately sent upstream.
The same issue affects routing. Two models may both advertise document input while supporting materially different file types, selection controls and billing rules. An AI API control plane for developer teams should represent those capabilities explicitly instead of treating “file input” as a single interchangeable feature.
What developers should check
Teams using Gemini should test representative Word, PowerPoint, Excel and OpenDocument files, including files with formulas, embedded media and complex layouts. They should compare whole-file and sliced requests for answer quality, usage and latency, while preserving selector metadata in logs.
Gateway operators have additional work: define a schema for the new controls, reject unsupported cross-provider routes, and update customer-facing metering to show which content was selected. The release can reduce conversion work and unnecessary processing, but only when those new request and billing semantics survive every layer between the application and Gemini.