Microsoft Copilot context window in 2026: 128K to 1M tokens by product

Updated: November 8, 2026

Microsoft does not publish a single context window for Microsoft Copilot. The figures it does document depend on the product: Copilot Studio lists 128K tokens for its GPT-4.1, GPT-5 chat and GPT-5.3 chat models, 200K tokens for Claude Sonnet 4.6 and Claude Opus 4.6, and 400K tokens for GPT-5 reasoning and GPT-5.2 reasoning, while GitHub Copilot offers an optional 1 million-token extended context on supported models in Visual Studio Code and Copilot CLI.

The consumer Copilot app and Microsoft 365 Copilot Chat have no published token figure at all. For those products Microsoft documents file and retrieval limits instead, such as 20 files of up to 50 MB per conversation, and a separate memory system that stores preferences across chats. The table below gathers every official number in one place, followed by a capacity estimate in words and pages, the memory controls for each product and the steps to check or extend context in practice.

··········

⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣

MICROSOFT COPILOT CONTEXT WINDOW BY PRODUCT.

The Copilot brand covers several products built on different models and retrieval systems, so the useful question is which Copilot you are using. The limits below come from Microsoft Learn, Microsoft Support and GitHub Docs as published on November 8, 2026.

........

Copilot product	Documented context or limit	What Microsoft publishes
Microsoft Copilot app (personal accounts)	No token figure published	20 files per conversation, 50 MB per file
Microsoft 365 Copilot Chat	No token figure published	Same upload limits; answers draw on retrieved work content
Microsoft 365 Copilot agents	No token figure published	Full-content search of up to 20 specified files, about 300 pages recommended
Copilot Studio	128K, 200K or 400K tokens per model call	Depends on the model selected for the prompt or agent
GitHub Copilot	Default size per model, optional 1M tokens	Extended context in VS Code and Copilot CLI on supported models

........

The consumer and Microsoft 365 products hide the model's raw window behind retrieval. When a question refers to a file, an email thread or a SharePoint site, Copilot searches the source and passes the relevant passages to the model, so a document can be far larger than what the model reads in one step. Copilot Studio and GitHub Copilot expose the model's window more directly because developers choose the model.

··········

COPILOT STUDIO TOKEN LIMITS FOR EACH MODEL.

Microsoft's Copilot Studio documentation on model settings, last updated on August 5, 2026, gives a context figure for each model available to prompts. Microsoft adds that the window applies to each individual model call, and that the standard agent harness does not enforce a token limit across the several calls a single conversational turn can make.

........

Copilot Studio model	Documented context	Licensing rate
GPT-4.1 mini (default)	Input up to 128K tokens	Basic
GPT-4.1	Up to 128K tokens	Standard
GPT-5 chat	Up to 128K tokens	Standard
GPT-5.3 chat	Up to 128K tokens	Standard
GPT-5 reasoning	Up to 400K tokens	Premium
GPT-5.2 reasoning	Up to 400K tokens	Premium
Claude Sonnet 4.6	Up to 200K tokens	Standard
Claude Opus 4.6	Up to 200K tokens	Premium
Grok 4.1 Fast (Non-reasoning)	Not stated	Standard

........

The Anthropic models are classed as experimental, and Microsoft describes Grok 4.1 Fast as not recommended for production use. Availability varies by region, and US government regions still use GPT-4o mini and GPT-4o in place of the GPT-4.1 models. The page lists no maximum output size for any model.

Agents built in Copilot Studio can also hold far more material than any window through knowledge sources. Microsoft's file-upload documentation allows up to 500 files per agent, each no larger than 512 MB, limited in total by the Dataverse storage of the environment. Those files are searched and retrieved in pieces, which is why an agent can answer from thousands of pages with a 128K model.

··········

GITHUB COPILOT 1 MILLION-TOKEN EXTENDED CONTEXT.

GitHub announced larger context windows and configurable reasoning levels on June 4, 2026. Its model reference now lets users choose between each model's default context size and an extended context of 1 million tokens, but only on the models marked as supporting it and only in Visual Studio Code and Copilot CLI (the June changelog also named the GitHub Copilot app). GitHub does not publish the default size for each model on that page.

The larger window has a cost. GitHub states that choosing a larger context window or a higher reasoning level increases AI credit consumption per interaction, and it recommends the defaults for everyday tasks. The extended option pays off on work that spans a large repository, long multi-file changes or long supporting documents.

GitHub also documents how sessions behave as the window fills. In Copilot CLI, automatic compaction starts at about 80% of the window's capacity and summarizes earlier turns; if usage reaches about 95% before compaction finishes, the CLI pauses until it completes. Compaction keeps goals, progress, important files and next steps, may drop fine-grained details, and cannot be reversed. Tool output larger than 20 KiB is saved to a temporary file by default, and the model receives the path and a preview.

··········

MICROSOFT 365 COPILOT DOCUMENT AND RETRIEVAL LIMITS.

For Microsoft 365, the practical ceiling is how much content Copilot retrieves and reads, and Microsoft documents it in characters, pages and file counts. Its guidance on optimizing content retrieval for agents gives these figures.

........

Microsoft 365 Copilot agent source	Documented limit
Files specified by the agent	Full contents searched for up to 20 files; above 20, the 20 most relevant
Total size of those files	Best practice of no more than 300 pages
SharePoint sites or folders by URL	Keep each file under 36,000 characters, about 15-20 pages
Embedded files	First 750-1,000 pages (1.8 million characters) of each file indexed
Excel content	Store the data in one sheet of the workbook

........

Microsoft also notes that Copilot cannot currently parse tables and other special formatting in SharePoint content. Permissions apply on top of every limit: Copilot only retrieves organizational data the signed-in user is allowed to open, and a larger model window does not change that.

For uploads in chat, Microsoft's support page applies the same limits to personal accounts, work or school accounts and free Copilot Chat: up to 20 files per conversation and up to 50 MB per file, in PDF, DOCX, XLSX, PPTX, common image formats, TXT, JSON, CSV and MD. Our guide to Microsoft Copilot file uploads and limits covers formats and reading behavior in more detail.

··········

[[ADPLUS_300x250_1]]

MICROSOFT COPILOT CONTEXT WINDOW IN WORDS AND PAGES.

To make the token figures easier to compare, Data Studios converted them with a common approximation of 0.75 English words per token and 500 words per page. These are raw capacity estimates; real sessions reserve part of the window for system instructions, tool definitions, retrieved passages, history and the answer itself.

........

Raw context	Approx. English words	Approx. 500-word pages	Where Microsoft documents it
128K tokens	~96,000	~192	Copilot Studio GPT-4.1 and GPT-5 chat models
200K tokens	~150,000	~300	Claude models in Copilot Studio
400K tokens	~300,000	~600	GPT-5 and GPT-5.2 reasoning in Copilot Studio
1M tokens	~750,000	~1,500	GitHub Copilot extended context

........

A 1 million-token window holds about 7.8 times the raw text of a 128K window and about 2.5 times that of a 400K window. Output quality does not grow in the same proportion, because answers depend on how relevant the supplied material is and how well the model uses long inputs.

··········

MICROSOFT COPILOT MEMORY AND CHAT HISTORY.

Context and memory are separate features. The context window is what the model reads during one answer; memory is a stored list of facts and preferences that Copilot adds to future conversations, and chat history is the record of past conversations you can reopen.

In the Microsoft Copilot app for personal accounts, an updated version for web, desktop and mobile has been available since August 18, 2026. In that version, saved memories are managed under Settings, Personalization, Saved memories, where you can delete single memories or all of them. Turning the Saved memories toggle off stops new memories but does not delete existing ones, so they must be removed separately. Past conversations stay available until you delete them from the chat list or from the Microsoft privacy dashboard. Microsoft's article for the older version of the app states an 18-month retention period for conversation history; the article for the new version gives no period.

Microsoft 365 Copilot has its own Copilot Memory for work accounts. Saving is on by default, Copilot may ask before storing something it considers important, and a request such as "remember that I prefer Excel for reporting" produces a "Memory updated" notice. Memories are managed in Copilot Chat under Settings and more, Chat settings, Personalization, Saved memories. Copilot Studio added memory for agents in preview on September 2, 2026, letting an agent retain information from previous interactions with an individual user when the maker enables it.

........

Copilot product	Memory feature	Where to manage it
Microsoft Copilot app (new version)	Saved memories	Settings, Personalization, Manage saved memories
Microsoft 365 Copilot Chat	Copilot Memory, on by default	Chat settings, Personalization, Saved memories
Copilot Studio agents	Agent memory (preview)	Enabled by the agent maker
GitHub Copilot	Session context with compaction	/context and /compact in Copilot CLI

........

··········

HOW TO CHECK AND EXTEND COPILOT CONTEXT IN PRACTICE.

In GitHub Copilot, the /context command in Copilot CLI and the GitHub Copilot app shows the model, the tokens in use and a breakdown across system prompt, custom instructions, tools, messages and free space. Supported IDEs show a context indicator next to the chat input, a small circle in Visual Studio Code. The /compact command starts compaction manually, and a fresh session is the better choice after switching to an unrelated task or after repeated compactions have worn away important details.

In Copilot Studio, the model choice sets the ceiling: a 128K model handles ordinary agent conversations, while GPT-5 reasoning or GPT-5.2 reasoning gives 400K tokens per call at the premium rate. Large reference material belongs in knowledge sources, where retrieval selects the relevant passages.

In Microsoft 365 Copilot and the consumer app, results improve when you attach fewer, more relevant files, split very long SharePoint documents into shorter ones and keep spreadsheet data in a single sheet. In long chats, restate key requirements, because earlier turns may be summarized or retrieved selectively even though they remain visible on screen. For the models behind each Copilot product, see our guide to Microsoft Copilot models in 2026, and for plan prices our overview of Microsoft Copilot pricing and free tier limits.

··········

⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁤

MICROSOFT COPILOT CONTEXT WINDOW: CONFIRMED FIGURES AND GAPS.

The confirmed numbers are product-specific. Copilot Studio documents 128K, 200K and 400K tokens per model call, GitHub Copilot documents an optional 1 million-token context in Visual Studio Code and Copilot CLI at a higher AI credit cost, and Microsoft 365 Copilot documents retrieval limits of 20 fully searched files, about 300 pages and 36,000 characters per SharePoint file.

The gaps are equally clear. Microsoft publishes no token figure for the consumer Copilot app or Microsoft 365 Copilot Chat, GitHub does not list default context sizes per model on its reference page, and the new consumer app's support article states no retention period for chat history. Readers who need a guaranteed window for long documents should use Copilot Studio with a model whose limit is documented, or GitHub Copilot's extended context for code.

··········

FOLLOW US FOR MORE.

·····

DATA STUDIOS

·····

datastudios.org