Microsoft Copilot Context Window Explained: token limits, memory behavior
Updated: Sep 15
Microsoft Copilot no longer has one meaningful context-window number because the Copilot name now covers several products that assemble information in different ways.
Microsoft 365 Copilot often retrieves relevant material from files, apps, and Microsoft Graph rather than loading every source directly into one raw model prompt, while Copilot Studio exposes model-level limits more explicitly and GitHub Copilot can use an extended 1 million-token context on supported models in Visual Studio Code and Copilot CLI.
The practical question is therefore not simply “How many tokens does Copilot have?” but how much information each Copilot surface can retrieve, place into the active working context, preserve during a long thread, and safely use under the permissions of the account or organization.
··········
COPILOT DOES NOT HAVE ONE UNIVERSAL CONTEXT WINDOW
A single context-window figure is misleading because Microsoft Copilot, Microsoft 365 Copilot, Copilot Studio, and GitHub Copilot can use different models, tools, retrieval systems, and product-specific limits.
In Microsoft 365, a user may ask about a Word file, a presentation, email, or organizational content and receive an answer assembled from retrieved sections rather than from a prompt containing every byte of every source. In Copilot Studio, model configuration exposes context limits directly. In GitHub Copilot, the available context depends on model and surface, and selected supported models can use an extended 1 million-token mode.
........
Copilot surface | Documented context behavior | Current practical implication |
Microsoft 365 Copilot | Retrieval, files, app context, and Microsoft Graph; no single public token number for every end-user surface | Large-source tasks depend on what Copilot retrieves, not only raw model capacity |
Copilot Studio | 128K for several chat/general models; 400K for documented GPT reasoning models; 200K for listed Claude models | Model choice directly changes available prompt capacity |
GitHub Copilot | Model-dependent default context plus optional 1M-token extended context on supported models | Very large repositories can use much more context in VS Code and Copilot CLI when enabled |
........
··········
MICROSOFT 365 COPILOT EXTENDS PRACTICAL CONTEXT THROUGH RETRIEVAL
Microsoft 365 Copilot is designed around retrieval. When a task references organizational content, the system can search relevant files and data, select useful sections, and provide those sections to the model instead of treating every source as one continuous document that must fit entirely inside a single model window.
A useful documented example comes from Microsoft 365 agent content retrieval: an agent can specify up to 20 files and search their full contents, while Microsoft recommends keeping the combined relevant material to no more than roughly 300 pages when the goal is reliable full-file search. That is a retrieval guideline for that agent scenario, not a universal 300-page limit for every Copilot product.
This distinction removes one of the biggest problems in the previous version of this article: a document can be much larger than the material actually sent to the language model in one generation step. Effective document reach and raw context-window size are related, but they are not the same measurement.
[[ADPLUS_300x250_1]]
··········
COPILOT STUDIO MAKES MODEL-LEVEL TOKEN LIMITS VISIBLE
Copilot Studio provides the clearest current example of why context capacity must be tied to the selected model rather than the Copilot brand itself.
........
Model in Copilot Studio | Documented context | Typical role |
GPT-4.1 / GPT-4.1 mini | 128K tokens | General and lightweight prompts |
GPT-5 / GPT-5.3 chat | 128K tokens | General chat workflows |
GPT-5 reasoning / GPT-5.2 reasoning | 400K tokens | Longer and deeper reasoning tasks |
Claude Sonnet 4.6 / Opus 4.6 | 200K tokens | External-model workflows |
........
The maximum theoretical window still is not identical to usable user input because system instructions, tool definitions, retrieved records, conversation history, and output generation consume part of the available budget.
··········
GITHUB COPILOT CAN REACH A 1 MILLION-TOKEN EXTENDED CONTEXT
GitHub Copilot has moved far beyond the old assumption that coding Copilot sessions sit somewhere around 64K or 128K tokens. GitHub now documents a 1 million-token extended context capability for supported models.
The extended mode is currently available in Visual Studio Code and Copilot CLI, and it is model-dependent. A user can choose the regular context for ordinary work or the 1 million-token mode when a task genuinely needs a much larger repository, long multi-file history, or extensive supporting material.
The larger window has an economic consequence: GitHub explicitly notes that longer context consumes more tokens and therefore more AI credits. Capacity is therefore a resource decision as well as a technical feature.
··········
CONTEXT WINDOW, RETRIEVAL, CHAT HISTORY, AND MEMORY ARE DIFFERENT LAYERS
A context window is the active working set available to the model for one generation. Retrieval is the mechanism that selects external material to place into that working set. Chat history is the stored conversation record shown to the user, and memory or personalization features are separate product systems that may preserve selected information across sessions.
This means that a conversation can remain visible in history while older details are no longer fully active in the current model context. Long-running agent and coding sessions may use summarization or compaction to reduce earlier material so the session can continue without sending the entire raw history on every turn.
In Microsoft 365, permissions add another layer: Copilot can only retrieve organizational data that the user is allowed to access. A larger context window does not bypass Microsoft 365 permissions, and broader retrieval does not imply permanent memory of every retrieved file.
··········
DATA STUDIOS NORMALIZED CAPACITY ESTIMATE SHOWS THE SCALE DIFFERENCE
To make the raw token figures easier to compare, Data Studios can normalize them using a simple approximation of 0.75 English words per token and 500 words per page. These figures are capacity estimates, not guarantees of usable document length, because real Copilot sessions reserve space for instructions, tools, retrieved context, conversation state, and output.
........
Raw context | Approx. English words | Approx. 500-word pages | Where this scale appears |
128K tokens | ~96,000 words | ~192 pages | Several Copilot Studio chat models |
200K tokens | ~150,000 words | ~300 pages | Listed Claude models in Copilot Studio |
400K tokens | ~300,000 words | ~600 pages | GPT reasoning models in Copilot Studio |
1M tokens | ~750,000 words | ~1,500 pages | Extended GitHub Copilot context on supported models |
........
A 1 million-token window therefore represents roughly 7.8 times the raw capacity of a 128K window, but the difference in real output quality will depend on retrieval quality, model behavior, instruction overhead, and how much of the supplied context is actually relevant to the task.
··········
LARGE FILES AND LONG THREADS NEED DIFFERENT WORKFLOWS
For Microsoft 365 work, supplying fewer relevant files is usually better than indiscriminately attaching an entire archive because retrieval quality becomes the constraint before theoretical token capacity does.
For Copilot Studio, the model should be selected according to the actual context requirement: a 128K model may be sufficient for ordinary agent interactions, while a 400K reasoning model gives substantially more room when the workflow genuinely needs a larger active prompt.
For GitHub Copilot, the 1 million-token option is most valuable when code has to be reasoned about across a large repository or long multi-file task. Using extended context for routine prompts can consume more AI credits without producing a proportional benefit.
For long conversations, users should assume that old material can be compacted, summarized, retrieved selectively, or fall outside the active working context even when the chat remains visible. Critical requirements should therefore be restated or kept in files and instructions that the active workflow can reliably retrieve.
··········
COPILOT CONTEXT IS NOW A PIPELINE PROBLEM, NOT A SINGLE TOKEN NUMBER
The most accurate way to evaluate Microsoft Copilot context in 2026 is to separate model capacity from retrieval, product surface, permissions, file handling, and session management. Microsoft 365 Copilot can reach beyond a raw prompt through retrieval; Copilot Studio exposes model windows that currently span 128K to 400K in documented Microsoft models; and GitHub Copilot can extend selected coding workflows to 1 million tokens.
That architecture also explains why memory, visible chat history, and context capacity should never be treated as synonyms. The quality of a Copilot workflow depends on which information is selected, how much enters the active context, what the model can process, and what the product preserves or retrieves on later turns.
··········
FOLLOW US FOR MORE
··········
DATA STUDIOS
··········
