What Is a Context Window? Claude Tokens and Context Limits by Plan — 2026 Guide

What Is a Token? The Unit Models Count In

Any discussion of the context window must begin with the token, because every figure published — 200K, 500K and 1M — is expressed in tokens. Not words, not characters, and not the number of messages a user may send.

A token is the smallest unit a language model operates on. It may be a complete word, part of a word, a single character, or a byte in the case of Unicode. Anthropic's glossary gives an estimation rule of approximately 3.5 English characters per token, while noting that the actual figure varies by language. The rule is therefore suitable for preliminary estimates, but not for budgeting or contractual commitments.

The figures published in Anthropic's documentation provide a more reliable reference point:

Content placed into context	Approximate tokens	Point of comparison
A typical web page (~10 kB)	~2,500 tokens	Several hundred fit within a 1M window
A long documentation page (~100 kB)	~25,000 tokens	Approximately eight fill a 200K window
A research paper in PDF (~500 kB)	~125,000 tokens	A single file occupies over half of a 200K window
A 200K-token context window	—	Roughly 500 pages of text or more, as stated in the help centre
Tool-use system prompt overhead (excluding the tool definitions themselves)	~286–804 tokens	Consumed before the conversation begins

# Measure actual token counts before sending content to the model
curl https://api.anthropic.com/v1/messages/count_tokens \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-opus-5",
    "messages": [
      {"role": "user", "content": "The text you want to measure..."}
    ]
  }'
# Response: {"input_tokens": 42}

What a Context Window Is: Working Space, Not Long-Term Memory

Anthropic's documentation defines the context window as all the text a model can reference while generating a response, including the response itself. The documentation describes it as the model's "working memory," and distinguishes it clearly from the large corpus used to train the model.

Put differently, knowledge acquired during training is what the model already carries; the context window is a bounded working space for the material relevant to the task at hand. Once that space is full, earlier content must be managed before new content can be added, and anything outside the space cannot be referenced.

A common misconception is that this space holds only what the user types. In practice, every component of the request is counted:

Component occupying the context window	Counted?	Why it is commonly overlooked
The system prompt	Yes	Not visible to the user, but already consuming space
Every message in the conversation, from both user and model	Yes	Earlier turns accumulate and are not discarded
Attachments, images and PDF files	Yes	A single large file can occupy six figures in tokens
Tool definitions and the results tools return	Yes	More integrations leave less initial working space
The model's reasoning process (extended thinking)	Yes	Deeper reasoning carries both a space and a cost implication
The response being generated	Yes	Space must be reserved for the answer, not filled entirely with input
Content stored through prompt caching	Yes	Caching reduces cost; it does not exempt content from being counted

Context Window by Claude Plan (as of 4 September 2026)

This is a question organisations evaluating a plan raise often, and a frequent source of confusion, because the accurate answer is that context window size is a property of the model, not of the plan being purchased. The plan determines which models can be accessed and how much they can be used; the figure of 200K, 500K or 1M derives from the model selected at that moment.

Plan	Context in chat	Context in Claude Code	Notes
Free	200K tokens	—	The pricing comparison table states 200k
Pro	Model-dependent, up to 1M	Up to 1M	Opus models require usage credits to be enabled
Max	Model-dependent, up to 1M	Up to 1M	Differs from Pro in quota, not in context
Team	Model-dependent, up to 1M	Up to 1M	Standard and Premium differ in usage volume
Enterprise	500k on the default model	Up to 1M	Wording as published on the pricing page

Each figure above derives from a model-level property. Anthropic's help centre states that when chatting with Claude on a paid plan, Fable 5.1, Opus 5 and Sonnet 5 support a 1M-token context window; Opus 4.8, Opus 4.7, Opus 4.6 and Sonnet 4.6 support 500K tokens; and remaining models operate at 200K tokens. On the API the position is more uniform, since the figure follows the model's own capability:

Model	Context window (via API)	Maximum output per request
Claude Fable 5.1 / Fable 5	1M tokens	128K tokens
Claude Opus 5 / 4.8 / 4.7 / 4.6	1M tokens	128K tokens
Claude Sonnet 5 / Sonnet 4.6	1M tokens	128K tokens
Claude Sonnet 4.5	200K tokens	—
Claude Haiku 4.5	200K tokens	—

These are the figures for API access; for chat, refer to the plan-level table above. The documentation specifies a 128K maximum output only for models with a 1M-token context window, and does not state a figure for the 200K models.

Three Common Misconceptions

1. A context window is not a usage limit

These are two different constraints. The context window determines how much a single conversation can hold; the usage limit determines how much can be used within a given period, calculated by Claude in rolling five-hour sessions with weekly ceilings applied on top. Moving from Pro to Max therefore increases quota rather than expanding context. Organisations that conflate the two frequently misjudge the effect of an upgrade. Quota details for each plan are set out in our comparison of every Claude plan's limits.

2. A token is not a word, and 1M does not hold the same content on every model

Anthropic's pricing documentation states that Claude 4.7 and later models use a new tokenizer that produces approximately 30% more tokens for the same text, noting that the actual increase depends on the content and the shape of the workload. Consequently, a 1M-token window on Opus 4.7 and above holds less content than a 1M-token window on Sonnet 4.6, even though the published figures are identical. Token volumes should therefore be re-measured whenever the model changes, rather than carried forward from earlier estimates. Further detail on generational changes is available in our coverage of the Claude Opus 5 launch.

3. A larger window does not necessarily produce a better answer

Anthropic's documentation states the point directly: "more context isn't automatically better." As the token count grows, accuracy and recall decline — a phenomenon the documentation refers to as context rot. Curating what enters the context is therefore no less important than the size of the space available. In practice, attaching a large volume of unfiltered documents tends to produce weaker results than attaching only those that are directly relevant.

When a Conversation Reaches Its Limit

Within the Claude applications, context is managed automatically as a conversation approaches its limit, with earlier messages summarised so that the conversation may continue. The documentation states that this process is not counted toward the usage limit, and that the full conversation history remains available for reference.

Behaviour on the API differs. Where input alone exceeds the context window, the request returns a "prompt is too long" error. Where generation reaches the limit mid-response, models from 4.5 onward halt with the status model_context_window_exceeded so that the application can handle the outcome. For long-running workloads, server-side compaction is recommended rather than allowing requests to reach the limit.

For business users, the following practices provide effective control over context:

• Begin a new conversation when the subject changes — unrelated history consumes space without improving the quality of answers.
• Hold standing reference material in Projects rather than re-attaching files in every conversation (see Claude Projects and Knowledge).
• Attach only the relevant sections — extracting the pages required from a large document yields better results than attaching it in full.
• Allow the system to retrieve data in response to the question rather than loading the entire dataset, via MCP or a retrieval process.
• Match the model to the nature of the task — a short summary does not require a 1M-token window.

Considerations for Organisations Running an ERP

A recurring pattern in organisations is the requirement for AI assistance in analysing a full year of procurement data, addressed by exporting the entire dataset to Excel and attaching it to a conversation. This approach carries the highest cost and the lowest accuracy of the available options, since a year of raw data consumes a substantial number of tokens and longer context reduces accuracy.

The more appropriate approach is to establish the ERP as the single source of reference and allow the assistant to retrieve only the records required to answer the question, through a connection with permission controls in place. This is precisely what the Model Context Protocol (MCP) was designed to provide.

Saeree ERP provides MCP integration services connecting an ERP system to Claude. The scope of the service comprises:

• Retrieval of live ERP records to answer questions — covering procurement, inventory, HR and document workflows. Users enquire in natural language and the system reads the records required to compose an answer, without exporting files into the conversation.
• Enforcement of the ERP's existing role-based permissions — users access data through the assistant within the same scope available to them in the system, with no exposure beyond their entitlements.
• An audit trail recorded for every data call — providing traceability of which user enquired about what, and which records were read to compose the answer.
• Operation with an ERP deployed on the organisation's own servers — data need not be copied elsewhere for the assistant to access it.

The practical outcome is that the context submitted to the model contains only the records relevant to the question, which is both lower in cost and higher in accuracy than loading a full year of data into the conversation.

Choosing Between Long Context and Targeted Retrieval

Nature of the task	Long context window	Retrieval or MCP
Analysing a single 200-page contract	Appropriate	Not required
Reviewing an entire repository in one pass	Appropriate	Not required
Querying the stock balance of a single item	Not appropriate	Appropriate — retrieve only the relevant record
Locating a regulation across 40 manuals	Not appropriate	Appropriate — search first, then submit the match
Daily tasks referencing the same document set	Not cost-effective	Appropriate — reduces both token cost and time

For teams still becoming familiar with the surrounding terminology — tokens, prompts, RAG and agents — our AI glossary for Thai enterprises consolidates the frequently used terms in one place.

"Effective context design is not measured by how much information can be loaded, but by how precisely the information necessary to the answer is selected."

- The Saeree ERP team

References

• Anthropic — Context windows (developer documentation)
• Anthropic Support — How large is the context window on paid Claude plans?
• Claude Pricing — plan comparison table
• Anthropic — Pricing (long context pricing and the tokenizer note)
• Anthropic — Glossary (Tokens and Context window definitions)
• Anthropic — Token counting API

All figures verified on 4 September 2026. Context window sizes and plan details change over time — confirm against the official sources before making a purchasing decision.