Claude Context Window Size: claude.ai Plans vs the API

You paste a long contract into Claude, ask three follow-up questions, and the chat starts "organizing its thoughts" or tells you it's out of room. The question you actually need answered is simple: how big is the Claude context window, and does your plan change it?

Short answer, as of October 2026: on paid claude.ai plans, the newest models get a 1M-token context window, some older ones get 500K, and anything not listed gets 200K. On the Claude API, every current model has 1M tokens, and older models such as Sonnet 4.5 have 200K. The model you pick matters more than the plan you pay for.

This post covers the Claude AI context window size by model and surface, how the API counts tokens against it, and what happens when you hit the limit. If you want the concept itself first, start with what a context window is. If you work in the terminal, the Claude Code 1 million context guide covers that surface in depth.

Claude AI context window size on claude.ai

Anthropic's Help Center article on context window sizes for paid plans lists the sizes per model. "Paid plans" means Pro, Max, Team and Enterprise. Here's the chat table as of October 2026:

For anything not on the list, the article says: "Outside of these models, Claude's context window size is 200K."

Notice what's missing: a per-plan column. Pro and Enterprise users picking the same model get the same window in chat. The plan decides which models you can pick and how much you can use in total. The model decides how much fits in one conversation.

What about the Free plan?

The Help Center table covers paid plans only. Claude's plans and pricing page lists the context window for every plan, Free included, as "Up to 1M" and "varies by model." The same page marks Sonnet and Haiku as available on Free, and Opus and Fable as paid-only.

Anthropic doesn't publish a separate per-model table for Free, so treat "up to 1M, depending on the model" as the honest answer for now.

Not all of the window is yours

Anthropic's article on usage and length limits adds a catch: "A portion of the context window is always reserved for Claude's response." The longest conversation you can have is a bit smaller than the headline number.

Context window in Cowork and Claude Code

The same model can get a different window depending on where you use it. The Help Center article keeps separate tables for chat, Claude Code and Cowork.

Claude Code is the most generous. Every model in the paid-plan table gets 1M tokens there. For Opus 4.6 and Sonnet 4.6 you select the 1M version explicitly, and the article says usage credits have to be enabled in some cases. The plan-by-plan rules, the [1m] variants and the ways you can end up stuck at 200K are covered in the Claude Code 1M context post linked above, so they aren't repeated here.

Cowork sits in between. As of October 2026, the Help Center lists:

The article also notes that Sonnet 5 automatically compacts a Cowork conversation at 500K tokens.

So the same Sonnet 4.6 conversation could hold 500K tokens in chat, 1M in Claude Code, and 200K in Cowork. If a task keeps running out of room, moving it to a newer model often helps more than moving it to a different app.

Claude context window on the API

The API rules live in Anthropic's context windows guide, and they're simpler than the app rules.

As of October 2026, these models have a 1M-token context window on the API: Fable 5.1, Fable 5, Opus 5.5, Opus 5, Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 5.5, Sonnet 5, Sonnet 4.6 and Haiku 5.5 (plus the Mythos models). Each can generate up to 128K output tokens in one request. Other models, including the deprecated Sonnet 4.5, have 200K.

Three details trip people up:

• No beta header needed. For every 1M model, 1M is the default. Older guides that tell you to add a long-context header are out of date.
• Long context is billed at standard rates, except on Haiku 5.5, where the docs say prompts over 100,000 tokens cost more. Anthropic's pricing page has the details.
• Images and PDFs hit a separate cap. One request can include up to 600 images or PDF pages, or 100 on 200K models. You can reach the request size limit before the token limit.

The models overview has the context window and max output for each model in one table. For the exact ID strings to put in your request, see the Claude model names reference.

How big is 1M tokens?

Anthropic's rough conversion: 1M tokens is about 555K words, or 2.5M characters, on the tokenizer introduced with Claude Opus 4.7. Older models fit about 750K words in 1M tokens. A 200K window holds roughly 150K words.

What counts toward the window

Everything in the request counts: the system prompt, every message (including tool results, images and documents), and your tool definitions. The output Claude writes for the turn counts too, thinking included.

Thinking is where the models differ. On Opus 4.5 and later Opus models, Sonnet 4.6 and later, Haiku 5.5, and the Fable and Mythos models, the API keeps earlier thinking blocks by default, and they count as input on later turns. On older models, the API strips earlier thinking blocks automatically.

Prompt caching changes the price of repeated tokens, not whether they count. Cached input still takes up space.

What happens when you go over

The docs describe two cases:

• The input alone is too big. Every model returns a 400 invalid_request_error with "prompt is too long."
• Input plus max_tokens is too big. On Claude 4.5 models and newer, the API accepts the request. If generation actually hits the limit, it stops with stop_reason: "model_context_window_exceeded". Older models return a validation error up front.

To avoid both, estimate a request with the token counting API before you send it. For long-running conversations, Anthropic points to server-side compaction, which summarizes earlier turns so the conversation can continue. It's in beta for Claude 4.6 and later models.

What happens when a claude.ai chat fills up

In the apps, running out of context doesn't have to end the conversation.

On paid plans with code execution enabled, Claude manages long chats automatically. When a chat nears the limit, it summarizes earlier messages to make room, and Anthropic says your full chat history is preserved so Claude can still reference it. That's what the "organizing its thoughts" message means.

Two caveats from Anthropic's Help Center articles on context windows and usage limits:

• Code execution has to be on. Without it, you hit the wall.
• "Longer conversations that trigger automatic context management consume more of your usage limit."

Edge cases still fail. Anthropic names very large first messages as one. If your first message is a giant file dump, split it up or put the material in a project.

Projects stretch the limit further

Projects get their own trick. Anthropic's article on RAG for projects says retrieval mode "automatically activates when your project approaches or exceeds the context window limits." Instead of loading every file into context, Claude searches the project's knowledge and pulls in the relevant parts. Anthropic says this lets you store up to 10x more content. It's available on Pro, Max, Team and Enterprise.

Context window vs usage limits

These two get mixed up constantly, and they fail in different ways.

• The context window (length limit) controls how long one conversation can get. Hit it, and that chat can't continue as-is.
• Usage limits control how much you can use Claude across all your conversations in a period. Hit them, and you wait for the limit to reset, unless you've enabled usage credits, in which case work continues on those.

A bigger plan raises your usage limits. It doesn't make a single chat longer if you're on the same model. If one long chat keeps hitting the wall, the fix is a model with a bigger window, automatic context management, or a project, not an upgrade.

How to fit more into Claude's context

Anthropic's own tips, plus a few that follow from how the window works:

1. Pick a 1M model. Opus 5.5, Sonnet 5.5 and Haiku 5.5 all get 1M in chat as of October 2026.
2. Turn on code execution so automatic context management can kick in.
3. Use projects for reference material. Only relevant content gets loaded.
4. Keep project instructions short. Put task-specific instructions in the chat instead.
5. Turn off what you don't need. Extended thinking, unused connectors and web search all add tokens.
6. Use a lower effort level for routine tasks, since higher effort uses more tokens.
7. Start a fresh chat when the topic changes. A long context also gets less accurate: Anthropic's docs call the effect "context rot."

FAQ

What is the Claude context window size?

As of October 2026, it's up to 1M tokens. On paid claude.ai plans, Fable 5.1, Opus 5.5, Opus 5, Sonnet 5.5, Sonnet 5 and Haiku 5.5 get 1M in chat. Several older models get 500K, and unlisted models get 200K.

Does Claude Pro have a bigger context window than Free?

Not by plan alone. Anthropic's pricing page lists "Up to 1M, varies by model" for every plan. The difference is which models you can use: Free includes Sonnet and Haiku, while Opus and Fable require a paid plan.

What is the context window for Claude on the API?

1M tokens for every current model and for most recent legacy models, with up to 128K output tokens per request. Other models, including Sonnet 4.5, Opus 4.5 and Haiku 4.5, have 200K.

Do I need a beta header for the 1M context window?

No. On every model with a 1M window, 1M is the default on the API.

Why does Claude say it's organizing its thoughts?

That's automatic context management. The chat is near its limit, so Claude is summarizing earlier messages to keep going. It needs code execution turned on, and it uses more of your usage limit.

Is Claude's context window the same as its memory?

No. The context window is what Claude can see in the current conversation. Memory across chats is a separate feature.

Key points

• Claude's context window depends on the model and the surface, not mainly the plan.
• In claude.ai chat, the newest models get 1M tokens, several older ones 500K, everything else 200K (as of October 2026).
• Claude Code gives 1M to every model in Anthropic's paid-plan table. Cowork varies from 200K to 1M.
• On the API, current models get 1M by default with no beta header. Sonnet 4.5, Opus 4.5, Haiku 4.5 and older models get 200K.
• Automatic context management and project RAG let long work continue past the window, at the cost of more usage.

A long, carefully built conversation is often the most valuable thing you made with an AI all week, and it's also the easiest to lose. Taku is an AI-native desktop workspace where you can save a working AI setup and run it again instead of rebuilding the context from scratch. Taku is in Beta, and the Mac app is available now.