# Claude Context Window Size: 200K to 1M Token Limits by Plan and Model

Claude's context window size is the maximum volume of combined prompt, conversation history, and uploaded file tokens that Anthropic's models can process in active working memory during a single session. On paid plans, Claude provides a 200,000-token capacity across Claude 3.5 and 3.7 Sonnet, scaling up to 1,000,000 tokens on expanded tiers. When document archives exceed active token memory, teams connect external workspaces over MCP for indexed retrieval.

Source: https://fast.io/resources/claude-context-window-size/
Author: [Tom Langridge](https://fast.io/authors/tom-langridge/)
Last reviewed: 2026-10-29

## How Large Is the Claude Context Window Across Models and Plans?

Anthropic's documentation ("How large is the context window on paid Claude plans?", verified October 2026) specifies that Claude's context window size is 200,000 tokens on baseline models (roughly 500 pages of text), while current flagship models scale to 1,000,000 tokens on paid subscriptions (Pro, Max, Team, and Enterprise).

Claude's context window size is the maximum volume of combined prompt, conversation history, and uploaded file tokens that Anthropic's models can process in active working memory during a single session. Unlike static training data frozen into model weights, the context window functions as active RAM. Every user message, system instruction, code snippet, attached spreadsheet, and previous assistant reply occupies space inside this working memory. Because transformer architectures evaluate attention across all tokens in the active buffer simultaneously, the context window establishes the practical ceiling for how much context the model can track before it forgets earlier instructions.

A 200,000 tokens standard context window across Claude 3.5 and 3.7 Sonnet serves as the primary operational baseline for daily chat, coding, and reasoning tasks. Tokens represent the fundamental units of text processed by language models. In standard English prose, one token corresponds to roughly three to four characters, translating to approximately 750 words per 1,000 tokens. Under this conversion, a 200,000-token context window holds approximately 150,000 words. That capacity fits a 500-page book, multi-chapter technical manuals, or dense financial audits in a single conversation. Frontier models supporting 1,000,000 tokens ingest roughly 750,000 words, enabling whole software repositories or dozens of research papers to sit directly inside working memory.

Context capacities differ across subscription tiers and operational environments. The table below outlines verified limits across Claude plans and product surfaces:

| Environment or Plan | Active Model Tier | Context Window (Tokens) | Document Capacity (Approx. Words) | Max File Upload Size | Source and Date |
| --- | --- | --- | --- | --- | --- |
| Claude Free (claude.ai) | Baseline Models | 200,000 tokens (session-throttled) | ~150,000 words | 500MB per file (chat, up to 20 files) | Verified October 2026 (support.claude.com) |
| Claude Pro / Team | Claude 3.5 Sonnet & Claude 3.7 Sonnet | 200,000 tokens | ~150,000 words (~500 pages) | 500MB chat / 30MB project files | Verified October 2026 (support.claude.com) |
| Claude Pro / Max / Team | Claude Opus 4.8 | 500,000 tokens | ~375,000 words | 500MB chat / 30MB project files | Verified October 2026 (support.claude.com) |
| Claude Enterprise & API | Flagship Models (Sonnet 5.5, Opus 5.5) | Up to 1,000,000 tokens | ~750,000 words | 500MB chat / 30MB project files | Verified October 2026 (support.claude.com) |
| Claude Code (CLI / Terminal) | Claude 3.7 Sonnet & Frontier Models | Up to 1,000,000 tokens (flag or credits) | Entire repositories (~750,000 words) | 30MB per repository file | Verified October 2026 (platform.claude.com) |
| Claude Cowork | Frontier Models | Up to 1,000,000 tokens (auto-compacts at 500K) | ~750,000 words | 30MB per project file | Verified October 2026 (support.claude.com) |
| Claude Messages API | Expanded API Tiers | Up to 1,000,000 tokens (model tier) | ~750,000 words | 30MB request payload | Verified October 2026 (platform.claude.com) |

While free Claude accounts provide access to the 200,000-token architecture, strict session limits restrict how many messages you can send before hitting cooldowns. Paid tiers (Pro, Team, and Enterprise) grant higher message volume allowances, project workspaces, and access to expanded 500,000 and 1,000,000-token models.

## What Counts Toward the Claude Token Capacity in Each Session?

Every interaction with Claude draws from a unified context budget. Many users assume that only their active prompt consumes tokens, but the context window must accommodate every piece of data relevant to the generation request. Understanding the seven primary components of the context budget helps teams avoid sudden context exhaustion:

1. System and Project Instructions: Foundational instructions define how Claude responds, its persona, and formatting constraints. In Claude Projects, project instructions sit as a persistent prefix across all project conversations, consuming tokens before you type your first character.

2. Conversation Transcript: Every message you send and every response Claude produces remains in memory for the duration of the thread. As a conversation lengthens, earlier messages consume an increasing proportion of the total token allowance.

3. Uploaded Text Files and Documents: Text extracted from TXT, Markdown, DOCX, HTML, RTF, EPUB, JSON, and source code files is tokenized directly. A 50-page technical specification containing 25,000 words translates to roughly 33,000 tokens.

4. Multimodal Images and Graphics: Image files (JPEG, PNG, GIF, WebP up to 8000x8000 pixels) convert into visual tokens based on pixel dimensions. Claude breaks images down into tiles, with high-resolution diagrams consuming several hundred to several thousand tokens each.

5. PDF Documents: Document parsing behavior depends on file size and page count. For documents submitted as PDFs, Anthropic notes that Claude analyzes both text and visual elements (like images, charts, and graphics) in PDFs of 100 pages or fewer, switching to text-only processing for documents between 101 and 1,000 pages. Anthropic documentation notes that PDFs exceeding 1,000 pages cannot be uploaded, triggering an error.

6. Tool Definitions and Connectors: External tools, custom connectors, and Model Context Protocol (MCP) integrations expose function schemas that explain their capabilities to the model. Every active connector consumes tokens upfront in the system prompt.

7. Extended Thinking Blocks and Model Output: When extended reasoning is active, the internal thinking tokens produced by the model during problem solving consume part of the generation budget alongside the visible response text.

Token density varies by document format. In standard English prose, one token corresponds to roughly three to four characters, yielding approximately 750 words per 1,000 tokens. In contrast, source code produces higher token density because programming languages rely on punctuation, camelCase naming conventions, indentation spaces, and syntax symbols that tokenizer dictionaries split into individual tokens. Structured spreadsheets (CSV and TSV) carry high token overhead: commas, quotation marks, and repeating column headers across thousands of rows quickly fill thousands of tokens. Raw JSON dumps and nested database records exhibit similar inflation due to repeated property keys.

## How Many Files Can You Put in a Claude Project?

A common myth across online technical forums asserts that Claude Projects enforces a hard ceiling of 50 or 100 uploaded files. That claim is incorrect.

Anthropic's official file upload documentation confirms that Claude Projects enforces no fixed file count cap. The documented rule establishes two criteria: an individual file upload ceiling of 30MB, and an unlimited number of files, provided the total content fits within Claude's context window.

The misconception about a 50-file limit arose from the mathematics of typical document lengths. Technical specifications, legal briefs, and architectural documents often span 10 to 15 pages, translating to roughly 3,000 to 4,000 tokens each. When users upload 50 of these documents into a project knowledge base running on the baseline 200,000-token context window, the accumulated text reaches 150,000 to 200,000 tokens. At that point, the visual capacity bar in the Claude Project sidebar fills completely, preventing further additions. If you upload concise markdown summaries of 500 tokens each, Claude Projects readily accepts hundreds of files without issue.

The operational rules for project knowledge differ from standard chat attachments:

* Direct Chat Conversations: Designed for immediate, temporary analysis. Chats accept up to 20 files per message, with individual files permitted up to 500MB. Once the conversation concludes, those files remain confined to that specific discussion thread.
* Claude Projects Knowledge: Designed for persistent reference across team workflows. Projects accept an unlimited number of files, but each individual file is capped at 30MB. Extracted text from all project files is indexed and referenced across every chat initiated within that project.

Exhausting the project capacity bar creates three operational bottlenecks:

First, conversational headroom contracts. Because project knowledge serves as a persistent background prompt, filling 180,000 tokens of a 200,000-token window leaves only 20,000 tokens for user questions, retrieved passages, thinking blocks, and Claude's answers. Long exchanges quickly run out of room.

Second, attention dilution impairs precision. When an LLM evaluates a prompt against an active context packed with hundreds of thousands of tokens, recall accuracy can suffer from the needle-in-a-haystack problem. Specific requirements buried in secondary project files may receive diminished attention.

Third, automatic RAG activates. To handle larger project knowledge bases on paid subscriptions (Pro, Max, Team, and Enterprise), Claude switches to automated retrieval-augmented generation (RAG) when project content nears context boundaries, loading relevant document chunks dynamically instead of placing all text into active memory.

## How Does Claude Manage Context Limits Automatically?

Managing long-running conversations requires strategies to prevent context overflow. On paid subscriptions with code execution enabled, Claude provides automatic context management.

When an ongoing conversation approaches the model's context window limit, Claude automatically summarizes earlier conversational turns to free memory for subsequent exchanges. This mechanism allows threads to continue without forcing the user to open a fresh conversation. During this process, you may occasionally observe a brief status notification indicating that Claude is organizing its thoughts.

Automatic context management preserves the high-level conversation record. Claude retains access to the overarching topics, factual decisions, and code changes discussed in earlier turns, even though verbatim tokens from older messages have been compressed. Anthropic notes that code execution must remain enabled for this automatic summarization to function; if disabled, long conversations hit standard token ceilings and prompt users to start a new chat.

While automatic summarization preserves thread continuity, packing massive volumes of information into active memory introduces a technical challenge known as context rot. Context rot describes the gradual degradation in model performance, reasoning precision, and instruction adherence that occurs as working memory accumulates conversational clutter, obsolete tool outputs, and tangential discussions.

Context rot stems from attention distribution in transformer architectures. As input token sequences expand toward 500,000 or 1,000,000 tokens, the model's attention mechanism must distribute its focus across millions of token pairings. Irrelevant data, outdated code versions, and verbose log traces compete with current user instructions for attention weight. Consequently, models operating in saturated context buffers may overlook subtle negative constraints, misinterpret variable declarations, or hallucinate details from superseded documents.

Prompt caching helps mitigate cost and latency for large contexts. Anthropic's prompt caching architecture stores the computational state of static token prefixes, such as project files, system prompts, and tool schemas. When consecutive turns share the same initial token sequence, Claude retrieves the precomputed attention state rather than reprocessing the entire document corpus from scratch. Prompt caching reduces response latency and lowers input token pricing. However, whenever you edit project instructions, add a file, or modify connected tools, the cache invalidates, requiring a fresh compilation pass.

## How Do You Query Large Document Collections Beyond Claude Context Limits?

Teams managing technical documentation, contract archives, research repositories, or client records eventually exceed Claude's active context window. Even with a 1,000,000-token capacity, a collection of 50 comprehensive PDF manuals, engineering schematics, and historical reports easily surpasses available memory.

Engineering teams address this boundary through three primary architectures:

1. Manual Project Partitioning: Teams divide files across multiple Claude Projects organized by functional domain. While simple, this approach creates information silos. Claude cannot reference files residing in Project B when answering a question inside Project A, forcing users to duplicate files and manage synchronizations manually.

2. Custom Retrieval Pipelines: Developers construct standalone retrieval-augmented generation systems using external vector databases, embedding models, and custom chunking scripts. While effective, custom RAG pipelines demand ongoing infrastructure maintenance, chunk tuning, and credential management.

3. External Workspaces Connected via Model Context Protocol: Rather than uploading raw files directly into Claude Projects or building private vector pipelines, teams store their corpus in an external workspace and connect Claude through an MCP server.

Fast.io provides an external workspace substrate designed for agentic teams and human collaboration, detailed in our [workspaces overview](/product/workspaces/) and [storage for agents guide](/storage-for-agents/). Instead of forcing entire document archives into Claude's active prompt tokens, files live in an organization-owned workspace. Fast.io leaves Claude's upload limits intact; what it adds is an indexed, searchable repository for files that do not fit into the context window.

When files enter a Fast.io workspace, Intelligence Mode automatically indexes documents for hybrid search, combining full-text keyword indexing, semantic meaning vectors, and search-by-metadata-value. When Claude needs information, it queries the workspace through the Fast.io remote MCP server, retrieving only the specific paragraphs, data rows, or citations required to answer the prompt. This architecture keeps Claude's active context window lean, eliminating context rot while granting the model access to gigabytes of source material.

Workspaces support direct file uploads or automated synchronization through Cloud Sync. Cloud Sync connects Google Drive, Dropbox, Box, and OneDrive folders to a workspace, one-way or two-way, on a schedule or on demand. Teams with unstructured files can also deploy [Metadata Views](/product/document-data-extraction/) to automatically extract typed structured schemas (dates, counterparties, totals, line items) from PDFs, spreadsheets, and scanned documents without manual OCR rules.

Connecting Claude to Fast.io takes minutes using the remote MCP server at `https://mcp.fast.io/mcp/tools`:

1. Open Claude on the web, desktop, or mobile app.
2. Open Customize, then select Connectors.
3. Add a custom connector and enter `https://mcp.fast.io/mcp/tools` as the URL.
4. Select Connect, and complete the OAuth sign-in to Fast.io in the browser window that opens.
5. In any chat conversation, turn Fastio on from the plus menu under Connectors.

Once connected, Claude reads and writes files according to member permissions configured per organization, workspace, and share. Collaborative Notes kept beside files allow human operators and AI agents to document research findings and instructions. For coding agents operating in terminal environments like Claude Code, Fast.io provides a dedicated code-mode endpoint at `https://mcp.fast.io/mcp/code`, configured by running `claude mcp add --transport http fast-io https://mcp.fast.io/mcp/code`.

Fast.io organization plans provide scalable capacity for engineering teams, as outlined on the [Fast.io pricing page](/pricing/):

| Plan | Monthly Price | Included Storage | Included Workspaces | AI Credits | Max Upload Size |
| :--- | :--- | :--- | :--- | :--- | :--- |
| Starter | $9.99/mo | 250 GB | 5 workspaces | 100,000 credits | 25 GB |
| Business | $49.99/mo | 5 TB | 50 workspaces | 600,000 credits | 50 GB |
| Enterprise | $199.99/mo | 25 TB | 200 workspaces | 3,000,000 credits | 100 GB |

Monthly plans start with a 14-day free trial that requires a credit card. Credits meter AI work. Storage beyond plan allowances is billed at 1.5 cents per GB a month, and extra bandwidth is billed at 4 cents per GB. Teams can review full setup instructions in the [MCP setup documentation](https://mcp.fast.io/docs).

## Frequently asked questions

### How big is the context window for Claude?

Claude provides a 200,000 tokens standard context window across Claude 3.5 and 3.7 Sonnet on paid plans, accommodating roughly 500 pages of text or 150,000 words. On expanded API and Enterprise tiers, flagship models support up to 1,000,000 tokens (approximately 750,000 words) for long-form reasoning, deep research, and repository-scale coding.

### What is the token limit for Claude Pro?

On the Claude Pro plan, token capacity depends on the active model. Claude 3.5 Sonnet and Claude 3.7 Sonnet provide a baseline 200,000 tokens, while models such as Claude Opus 4.8 provide 500,000 tokens, and flagship tiers provide up to 1,000,000 tokens. Claude Pro subscribers also receive automatic context management when code execution is enabled.

### How many files can I put in a Claude Project?

Claude Projects enforces no fixed numerical cap on uploaded files. You can upload an unlimited number of files, provided each individual file does not exceed 30MB and the cumulative text fits within Claude's context window. Early assumptions of a 50-file ceiling stemmed from typical technical documents exhausting the 200,000-token capacity.

### What is the difference between Claude chat file limits and Claude Project file limits?

Direct Claude chats accept up to 20 files per conversation at up to 500MB each, intended for temporary discussion within a single thread. Claude Projects limits individual files to 30MB but allows an unlimited file count, storing extracted text as persistent project knowledge referenced across every conversation in that workspace.

### What happens when Claude reaches its context window limit?

In standard conversations without automatic management, reaching the context limit produces a notification indicating that the conversation is too long, requiring you to start a new chat. On paid plans with code execution enabled, Claude automatically summarizes earlier conversation turns to free token space while retaining access to the full discussion record.

### How can I query documents that exceed Claude context window limits?

For document collections that exceed 200,000 or 1,000,000 tokens, teams store files in external workspaces like Fast.io and connect them to Claude through the Model Context Protocol at `https://mcp.fast.io/mcp/tools`. Fast.io indexes files automatically for hybrid search, allowing Claude to retrieve only the relevant passages on demand rather than loading entire document libraries into working memory.

## Sources

- [Claude Help Center: How large is the context window on paid Claude plans?](https://support.claude.com/en/articles/8606394-how-large-is-the-context-window-on-paid-claude-plans): Outside of expanded models, Claude standard context window on paid plans defaults to 200,000 tokens for baseline models (about 150,000 words), scaling to 1,000,000 tokens on flagship tiers and ingesting roughly 500 pages of text.
- [Claude Help Center: Upload files to Claude](https://support.claude.com/en/articles/8241126-upload-files-to-claude): Anthropic restricts individual files uploaded to Claude Projects to 30MB each, while direct chat conversations accept up to 20 files at up to 500MB each.

## About Fast.io

Fast.io provides shared workspaces where people and AI agents work on the same files, with built-in semantic search and citation-backed chat over what they hold. Agents reach it through a remote MCP server, a REST API at https://api.fast.io/current/, and a command line client published on npm as @vividengine/fastio-cli. MCP setup is at https://mcp.fast.io/docs: Claude and most MCP clients connect to https://mcp.fast.io/mcp/tools, ChatGPT to https://mcp.fast.io/mcp/operations, and coding agents to https://mcp.fast.io/mcp/code.