How Long Can ChatGPT Prompts Be? 2026 Limits and File Guide

There is no single published character cap for a ChatGPT message. The practical ceilings are the model's context window and a few documented thresholds: pastes over 10,000 characters turn into an attachment, a February 20, 2026 release note specifies 256,000 total tokens for manually selected Thinking (128,000 input, 128,000 maximum output), without establishing current Sol/Astra allowances, Enterprise file processing places up to 110,000 tokens of document text in a 128,000-token context, and the API models expose 1,050,000-token windows. Roughly, 1,000 tokens is about 750 English words. ChatGPT release notes, Enterprise file processing, Token guidance.

The limits OpenAI actually publishes

Distinguish three things: which models a plan can use, how much context a model can work with in one request, and the usage allowance that meters how often you can use it. None of them is a per-message character count, and OpenAI's current model-access pages do not list a context window for every plan and model combination. What is documented:

Sources: Thinking and large-paste release notes, Enterprise file processing, Custom Instructions FAQ, File Uploads FAQ, API model comparison.

Which models each plan gets

Model access comes from OpenAI's ChatGPT model guide, Business limits, Enterprise limits and the Business rename FAQ. The Thinking and Enterprise figures above are not mapped by OpenAI to each Sol or Astra option, so treat them as the documented reference points rather than a per-plan guarantee.

Plus access to Astra in Work and Codex does not establish an Astra context allowance in ordinary Chat. Use the GPT-6 Astra explainer for product context, and the model release timeline to navigate model names. Official access details.

Tokens vs characters vs words

Tokens are chunks of text: they can represent punctuation, characters, word fragments, or whole words. OpenAI gives these approximate relationships for English:

The larger rows apply OpenAI's rule of thumb arithmetically. They are estimates, not guaranteed conversions; language, spelling, spaces, capitalization, model, and encoding affect tokenization. Understanding and counting tokens.

The text you paste is only part of the budget. In the API, input, output, and reasoning tokens share context capacity. Earlier conversation turns and other supplied material also matter, so a context-window figure should not be treated as a fresh-message allowance. Conversation-state documentation.

For practical planning, count the material you actually intend to submit and leave room for the response. A plain-text count also does not capture all request overhead, such as message structure, tools, or images. OpenAI links its tokenizer and a complete Responses input-counting facility from its token guidance.

What happens when you exceed the limit

Different limits produce different situations. A large paste may become an attachment automatically; OpenAI documents a Show in text field control to convert it back. That action does not establish that the full text fits in the model's context. Large-paste behavior.

If ChatGPT rejects an input for size, shorten it or process the source in sections. The current help pages cited here do not publish a universal overflow rule specifying when ordinary ChatGPT rejects, truncates, or summarizes a conversation. Do not assume silent truncation always occurs, or that everything visible in a long transcript reaches every response.

Memory is a separate personalization feature that can draw useful context from chats, files, and connected apps. It is not evidence that an oversized prompt was processed in full. Keep source documents and explicit task requirements available rather than treating remembered context as a verbatim archive. Memory FAQ.

API overflow has documented controls, explained below; those controls should not be assumed to describe the ChatGPT interface.

How to fit a long prompt

Divide the task into meaningful sections

Split a long source at document headings or other natural boundaries. Give each section an identifier, specify the same extraction format, and combine the results afterward.

For example, ask each pass to return claims, supporting passages, unresolved questions, and source locations. Review those intermediate results before requesting a combined answer. Sending chunks into the same growing conversation should not be treated as an unlimited-context workaround.

Upload files and respect their separate limits

OpenAI's File Uploads FAQ publishes these restrictions:

An accepted file is not a promise that all its contents occupy the context window simultaneously. OpenAI's Enterprise guide describes combining directly included text with material retrieved from a search index. For document questions, specify the relevant section and ask for source locations so you can check coverage. Enterprise retrieval behavior.

Use projects for recurring work

Projects group files, conversations, and instructions. Project instructions apply inside that project and override global custom instructions. Use them for a stable objective, terminology, and output requirements. Projects documentation.

The file FAQ lists project capacities of 5 files for Free, 25 for Plus, and 40 for Pro, Business, and Enterprise. These are file counts, not context-window sizes. Project file limits.

Summarize with a verification step

Create a compact working summary containing decisions, constraints, unresolved issues, and pointers to the original material. Check it before continuing the task. Preserve exact wording only where the wording matters, and return to the source when a summary does not contain enough evidence.

For a large document collection, try summaries of individual documents before requesting a synthesis. This follows OpenAI's recommendation to narrow document questions and combine summaries for extensive material. File-analysis guidance.

Keep system-style instructions concise

Separate the task, constraints, requested format, and reference material. Use a short reusable template from the prompt library, then adapt it to the actual question.

For persistent preferences, custom instructions have their own documented character limits: 1,500 for Free and Go, and 5,000 for Plus, Pro, Business, Enterprise, and Education. These are custom-instruction limits, not message limits. Naming a section “system instructions” does not create extra context capacity. Custom Instructions FAQ.

API limits

API model specifications are separate from ChatGPT subscription allowances. OpenAI publishes the following context and output capacities:

Sources: Astra, Sol, and Terra comparison, Luna model documentation. Output capacity sits within the shared context budget; it is not extra input space. Context accounting.

The Responses reference marks truncation as deprecated and documents disabled as its default: an input exceeding context capacity fails with HTTP 400. With auto, items are dropped from the beginning of the conversation to fit. Choose deliberately whether discarding earlier context is acceptable. Responses reference.

For application budgeting, consult the LLM API pricing comparison. A published context maximum is a capacity specification, not a statement about ChatGPT plan access.

FAQ

What is the ChatGPT prompt character limit?

OpenAI publishes no universal per-message character cap. Pastes over 10,000 characters become an attachment, and custom instructions are capped at 1,500 or 5,000 characters depending on plan. The real ceiling is the model's token context, shared with the reply.

How many words can I put in a ChatGPT prompt?

OpenAI estimates approximately 75 English words per 100 tokens. That helps estimate text size, but it cannot determine your remaining prompt allowance without the applicable context budget and other included material.

Does uploading a PDF bypass the token limit?

No. Upload acceptance and model context are separate constraints. The file can meet upload limits while only selected material is used for a particular response; check the cited passages and request focused analysis.

Does ChatGPT Plus include a million-token prompt window?

That is not established by the current ChatGPT model-access documentation. The API's published 1,050,000-token windows should not be presented as Plus message limits.

Sources

• GPT-5.6 and GPT-6 Pro in ChatGPT
• ChatGPT Business models and limits
• ChatGPT Enterprise and Edu models and limits
• ChatGPT Business Rename FAQ
• ChatGPT release notes: Thinking context and large pastes
• Understanding and counting tokens
• File Uploads FAQ
• Optimizing File Uploads in ChatGPT Enterprise
• Memory FAQ
• Projects in ChatGPT
• ChatGPT Custom Instructions
• API model comparison: Astra, Sol, and Terra
• GPT-5.6 Luna API model
• API conversation state and context accounting
• Responses API reference: truncation behavior