Skip to content

Why Dumping a File Rarely Works

Why "analyze this" is not one request

As What AI Is Actually Doing With Your Data covers, the same instruction can route to genuinely different processing paths. Vendors do not even agree on what a given path guarantees, or on how large a request that path can even take. The three major chat tools do not share one context-window size. Anthropic's flagship consumer plan tops out at 200,000 tokens (Anthropic, 2026). OpenAI's current flagship model is documented at 1,050,000 tokens (OpenAI, 2026). Google's Gemini app scales from 32,000 tokens on its free tier up to 1,000,000 on its top paid tiers (Google, 2026). Size is only half the difference. Anthropic documents its code-execution tool as an explicit workaround for an oversized file. That tool moves the work out of the context window entirely, into a sandboxed environment (Anthropic, 2026). Google documents the opposite architecture for the same feature. Its code-execution mode is capped by the same token window as plain chat. A file too large for that window returns an error rather than routing around it (Google, 2026). Two vendors built the same-sounding feature to solve different problems. Neither failure looks different from the outside. Knowing which architecture, and which size limit, you are actually using matters before you trust either one.

What happens when a file does not fit

The three major vendors document three different answers to the same question. None of them is "the file gets fully processed regardless of size." Anthropic's consumer product, Claude.ai, automatically summarizes earlier parts of a long conversation once it nears the context window. That happens only when code execution is turned on. Without that setting, the same conversation can simply hit a limit (Anthropic, 2026). At the API level, Anthropic instead returns a named, explicit error. It reads "Exceeds context window size," and fires when a file is larger than the model can hold (Anthropic, 2026). Google's documented behavior is softer and harder to notice. Exceeding the context window does not error. It degrades, producing answers that miss connections or details spread across the content (Google, 2026).

The sharpest example of a silent, easy-to-miss limit sits with OpenAI, though it applies to one specific surface. It is the input_file parameter in OpenAI's developer-facing API, not the Advanced Data Analysis tool inside the ChatGPT app. That API does not read an entire spreadsheet once a sheet exceeds one thousand rows. Instead, it parses only the first one thousand rows directly. It substitutes a model-generated summary and header metadata for the rest of the sheet (OpenAI, 2026). A ten-thousand-row spreadsheet and an eleven-hundred-row spreadsheet come back in exactly the same answer format. Nothing marks which one was actually read in full. OpenAI does not publish a comparable row-level limit for the ChatGPT app's own data-analysis mode. Treat that surface as undocumented, not as safe. The three vendor behaviors above differ. They share one trait: each is silent by default, and a reader has to already know to look for it. Preparing Your Data Before You Ask covers what to check before you rely on any of them.

The formats that add a second layer of risk

File format adds a failure mode on top of size. The three common formats do not carry that risk equally. Anthropic's own Files API does not treat Excel files as a natively supported document type at all. A spreadsheet must first be converted to plain text, or routed through the code-execution tool, before Claude can work with it (Anthropic, 2026). Once a spreadsheet does reach a tool's code-execution mode, merged cells are a separately documented failure. Users have reported it directly on OpenAI's own developer community forum. The parsing library has no concept of a merged region. Values shift into the wrong cells without any error (OpenAI Developer Community, 2026).

PDF tables carry a different risk entirely, because a PDF has no structured cell data to begin with. Claude processes PDF pages as images. A dense page of small text, or a complex table, can exhaust the context window before the page limit is reached (Anthropic, 2026). A plain CSV file avoids both of these specific problems. It has no merge regions to misread and no page images to parse, since it is already a flat, structured text format. That does not make a CSV immune to the size limits covered above. It does remove an entire category of failure that Excel and PDF both add on top of size.

Sources

  • Anthropic — How do usage and length limits work?, 2026. Claude's context window size is 200K tokens across all models and paid plans, except for Enterprise plans, which have a 500K context window on some models. When your conversation approaches the context window limit, Claude summarizes earlier messages to continue the conversation seamlessly. Code execution must be enabled for automatic context management. View source · verified 2026-07-06 · primary
  • OpenAI — GPT-5.5 Model, 2026. 1,050,000 context window. 128,000 max output tokens. View source · verified 2026-07-06 · primary
  • Google — Gemini Apps limits & upgrades for Google AI subscribers, 2026. Without an AI plan: 32k tokens. AI Plus: 128k tokens. AI Pro & AI Ultra: 1 million tokens. View source · verified 2026-07-06 · primary
  • Anthropic — Files API, 2026. For file types that the document block doesn't support (for example, .docx and .xlsx), convert the files to plain text and include the content directly in your message... To analyze datasets instead of reading them as text, upload them for the code execution tool using a container_upload block. View source · verified 2026-07-06 · primary
  • Google — Code execution, 2026. The maximum file input size is limited by the model token window. If you upload a file that exceeds the model's maximum context window, the API will return an error. View source · verified 2026-07-06 · primary
  • Anthropic — Files API, 2026. Exceeds context window size (400): The file is larger than the context window size (for example, using a 500 MB plain text file in a /v1/messages request). View source · verified 2026-07-06 · primary
  • Google — Gemini Apps limits & upgrades for Google AI subscribers, 2026. If you exceed the context window, this could lead to responses that don't take into account all the content provided or miss connections or details throughout the content. View source · verified 2026-07-06 · primary
  • OpenAI — File inputs, 2026. For spreadsheet-like files (such as .xlsx, .xls, .csv, .tsv, and .iif), input_file uses a spreadsheet-specific augmentation process. Instead of passing entire sheets to the model, the API parses up to the first 1,000 rows per sheet and adds model-generated summary and header metadata so the model can work from a smaller, structured view of the data. View source · verified 2026-07-06 · primary
  • OpenAI Developer Community — Advanced Data Analysis GPT-4 | Excel Files - Bugs, 2026. cells move out, chatgpt is mistaken when there are merged cells in the file, it does not perceive them. View source · verified 2026-07-06 · ⚠ secondary mirror
  • Anthropic — PDF support, 2026. Dense PDFs (many small-font pages, complex tables, or heavy graphics) can fill the context window before reaching the page limit. Requests with large PDFs can also fail before reaching the page limit, even when using the Files API. View source · verified 2026-07-06 · primary