top of page

Claude context window in 2026: 1M tokens, 128K output and memory rules

Aug 11, 2025
9 min read

Updated: 3 days ago

Claude context window in 2026: 1M tokens, 128K output and memory rules, Data Studios original cover

Updated: November 15, 2026


Every current Claude model has a 1-million-token context window and can write up to 128,000 output tokens per response in the API. That covers Claude Fable 5.2, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5, the four models Anthropic lists as its latest lineup, and the same 1M figure applies to Claude Opus 5, Claude Sonnet 5, Claude Fable 5.1 and the earlier 4.6, 4.7 and 4.8 generations in the API. Older models such as Claude Opus 4.5, Claude Sonnet 4.5 and Claude Haiku 4.5 stay at 200,000 tokens.


In the Claude apps the figure depends on the model you pick. Anthropic's help center lists 1M tokens in chat for Fable 5.2, Fable 5.1, Opus 5.5, Opus 5, Sonnet 5.5, Sonnet 5 and Haiku 5.5 on paid plans, 500K tokens for Fable 5, Opus 4.8, Opus 4.7, Opus 4.6 and Sonnet 4.6, and 200K tokens for other models. The tables below give the window per model and per app, the long-context API prices, a cost example for a 600,000-token prompt, and the rules for memory across chats.


··········


⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣

CLAUDE CONTEXT WINDOW AND MAX OUTPUT BY MODEL.


Anthropic's models overview gives the context window and the maximum output of every model still listed in the API. The context window is the total number of tokens Claude can consider in one request, including the reply it writes. The max output is the cap on the reply itself, set with the max_tokens parameter.


........


Model

Context window

Max output

Input / output price per 1M tokens

Claude Fable 5.2

1M tokens

128K tokens

$8 / $40

Claude Opus 5.5

1M tokens

128K tokens

$4 / $20

Claude Sonnet 5.5

1M tokens

128K tokens

$2 / $10

Claude Haiku 5.5

1M tokens

128K tokens

from $0.10 / from $0.50

Claude Opus 5

1M tokens

128K tokens

$5 / $25

Claude Sonnet 5

1M tokens

128K tokens

$2 / $10

Claude Opus 4.8, 4.7, 4.6

1M tokens

128K tokens

$5 / $25

Claude Sonnet 4.6

1M tokens

128K tokens

$3 / $15

Claude Opus 4.5

200K tokens

64K tokens

$5 / $25

Claude Sonnet 4.5

200K tokens

64K tokens

$3 / $15

Claude Haiku 4.5

200K tokens

64K tokens

$1 / $5


........


Claude Mythos 5.2, Mythos 5.1 and Mythos 5 also have a 1M-token window and 128K output, but Anthropic offers them only to organizations verified through its access programs. On the Message Batches API, Opus 5.5, Opus 5, Sonnet 5.5, Sonnet 5, Haiku 5.5, Opus 4.8, Opus 4.7, Opus 4.6 and Sonnet 4.6 can write up to 300,000 output tokens with the output-300k-2026-03-24 beta header.


Anthropic converts the window into words for readers who think in pages. On the tokenizer introduced with Claude Opus 4.7 and used by the current models, 1M tokens is roughly 555,000 words or 2.5 million characters. Models before Opus 4.7 fit about 750,000 words into 1M tokens, and 200,000 tokens is roughly 150,000 words.


··········


CLAUDE CONTEXT WINDOW IN THE APPS: CHAT, CLAUDE CODE AND COWORK.


The Claude apps do not always expose the full API window. Anthropic's help article on paid plans gives separate lists for chat, Claude Code and Claude Cowork, and it applies to the Pro, Max, Ultra, Team and Enterprise plans without a different figure per plan.


........


Model

Claude chat

Claude Code

Claude Cowork

Fable 5.2, Fable 5.1

1M tokens

1M tokens

1M tokens

Opus 5.5, Opus 5

1M tokens

1M tokens

1M tokens

Sonnet 5.5

1M tokens

1M tokens

1M tokens

Sonnet 5

1M tokens

1M tokens

1M tokens (compacts at 500K)

Haiku 5.5

1M tokens

1M tokens

500K tokens

Fable 5, Opus 4.8, Opus 4.7

500K tokens

1M tokens

1M tokens

Opus 4.6, Sonnet 4.6

500K tokens

1M with [1m] variant

200K tokens

Other models

200K tokens

not listed

200K (Haiku 4.5)


........


In Claude Code, Opus 4.6 and Sonnet 4.6 reach 1M tokens only when the user selects the [1m] variant of the model with the /model command, and Pro subscribers need usage credits turned on for it. Anthropic does not publish a context window figure for the Free plan in the same article, so Free users should treat the paid-plan numbers as an upper bound.


A part of the window is reserved for Claude's reply, so the longest conversation you can hold in chat is slightly shorter than the headline figure. Files, images and project documents placed in the conversation use the same window as your messages. Our guide to Claude file upload limits covers the per-file and per-chat caps that apply before the context window is reached.


··········


WHAT COUNTS TOWARD THE CLAUDE CONTEXT WINDOW.


Anthropic's context window documentation lists everything that occupies the window in an API request: the system prompt, every message in the conversation including tool results, images and documents, the tool definitions, and the output of the current turn, thinking included. Each new turn sends the full earlier history again, so the reply of one turn is part of the input for the next.


Thinking tokens are part of max_tokens and are billed as output. On Claude Opus 4.5 and later Opus models, Sonnet 4.6 and later Sonnet models, Haiku 5.5 and the Fable and Mythos 5.x models, thinking blocks from earlier turns are kept by default and count as input in later requests. On earlier Opus and Sonnet models and on Haiku models up to Haiku 4.5, the API strips them automatically. Developers can change either default with thinking block clearing.


Prompt caching lowers the price of repeated input, but cached tokens still occupy the window. Anthropic's documentation states that cache reads and cache writes count toward the limit in the same way as uncached input. The documentation also warns about context rot, the gradual loss of accuracy and recall as a prompt grows, and recommends curating what goes into a long request rather than filling the window.


On models with a 1M-token window a single request can carry up to 600 images or PDF pages, against 100 on 200K-token models, although the request size limit may apply first.


··········


WHAT HAPPENS WHEN A CLAUDE CONVERSATION HITS THE TOKEN LIMIT.


In the Claude apps, Anthropic's help center says that for users with code execution enabled Claude now manages long conversations automatically. As the limit approaches, Claude summarizes earlier messages so the chat can continue, the full history stays visible, and Claude can still refer to it after the summary. A short "organizing its thoughts" indicator shows when this is running. These long chats draw more from the plan's usage limit, and a very large first message can still exceed the window.


Without code execution, or when a chat still runs out of room, the practical options are starting a new conversation or moving material into a project. Length limits are separate from usage limits: the length limit caps one conversation, while the usage limit caps how much you can use Claude across all chats, with a session limit that resets every five hours and weekly limits on paid plans. Our guide to Claude plans and usage limits explains those caps plan by plan.


In the API the behavior depends on the case. A prompt that by itself exceeds the window returns a 400 invalid_request_error with the message "prompt is too long" on every model. When the prompt fits but prompt plus max_tokens exceeds the window, Claude 4.5 and newer models accept the request and stop with stop_reason: "model_context_window_exceeded" if generation reaches the limit, while earlier models return a validation error unless the model-context-window-exceeded-2025-08-26 beta header is set. Developers moving between providers can check our guide to the right max_tokens parameter for OpenAI, Claude and Gemini.


··········


[[ADPLUS_300x250_1]]

CLAUDE MEMORY ACROSS CHATS AND PROJECTS.


The context window is memory inside one conversation. Memory across conversations is a separate feature, and Anthropic's help center describes it as follows. Memory is on by default for Free, Pro, Max and Ultra users on the web, desktop and mobile apps. Claude saves individual topics as you chat, instead of summarizing a conversation after it ends, and you can view, edit or delete those topics under Settings, Your Claude, Manage next to "Your memory", or ask Claude in chat to remember or forget something.


Chat search, which lets Claude look up earlier conversations, is listed for the Pro, Max, Ultra, Team and Enterprise plans and has its own toggle, "Search and reference chats". On Team and Enterprise plans the owner decides whether memory is available. In organizations created on or after October 31, 2026 memory starts on for members unless they turned it off before, while in older organizations it starts off until members enable it.


Each project has its own memory. Claude uses the memory of the current project plus your account memory, never the memory of other projects, and a project can be isolated by turning off "Use account memory" on its page. Incognito chats, available on every plan, are not saved to memory or chat history. Sensitive topics such as health, politics and religion are not saved unless you opt in, and deleting a conversation does not remove the memory topics created from it.


Projects also extend how much material Claude can work with. When project knowledge approaches or exceeds the context window, Claude switches to retrieval and searches the uploaded files for the passages relevant to each question. Anthropic says this lets paid plans store up to 10 times more content in a project, and the feature is available on Pro, Max, Ultra, Team and Enterprise. Retrieval brings selected passages into the window; the full knowledge base is never in context at once.


··········


CLAUDE API LONG-CONTEXT PRICING AND A 600,000-TOKEN COST EXAMPLE.


Anthropic's pricing page states that Claude 4.6 and later models, with the exception of Haiku 5.5, include the full 1M-token window at standard prices, so a 900,000-token request costs the same per token as a 9,000-token request. Claude Haiku 5.5 is tiered: when the prompt is over 100,000 tokens, every token in the request is billed at the higher rate, and the count includes cache reads and cache writes.


........


Claude Haiku 5.5 price per 1M tokens

Prompt up to 100K tokens

Prompt over 100K tokens

Base input

$0.10

$0.50

Base output

$0.50

$2.50

Cache hit

$0.01

$0.05

Batch input

$0.05

$0.25

Batch output

$0.25

$1.25


........


The example below assumes one request with a 600,000-token prompt, such as a large contract set or a code repository, and a 4,000-token answer, at the standard prices listed on Anthropic's pricing page on November 15, 2026. The Batch API halves both input and output prices.


........


Model

Cost per request

With Batch API

Claude Fable 5.2

$4.96

$2.48

Claude Opus 5.5

$2.48

$1.24

Claude Sonnet 5.5

$1.24

$0.62

Claude Haiku 5.5 (over-100K rates)

$0.31

$0.155


........


Prompt caching changes the result for repeated questions over the same documents. A cache read costs $0.20 per million tokens on Opus 5.5 and $0.10 on Sonnet 5.5, so a follow-up question that reads the same 600,000 tokens from cache costs about $0.20 on Opus 5.5 and $0.10 on Sonnet 5.5 with the same 4,000-token answer, before the one-time cache write. On Haiku 5.5 the tier boundary carries weight: a 90,000-token prompt with a 4,000-token answer costs about $0.011, while a 110,000-token prompt costs about $0.065 because the whole request moves to the higher rates.


··········


⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁤

HOW TO CHECK AND MANAGE CLAUDE TOKEN USAGE.


Developers can read the limits of each model from the Models API, whose response includes max_input_tokens, max_tokens and a capabilities object, and can estimate a request before sending it with the token counting endpoint. Every response reports actual consumption in the usage field, split into input_tokens, cache_read_input_tokens and cache_creation_input_tokens when caching is used.


For long agent sessions Anthropic points to server-side compaction, which summarizes earlier parts of the conversation on the server so work can continue past the window, in beta for Claude 4.6 and later models. Context editing adds two narrower tools, tool result clearing for old tool outputs and thinking block clearing for earlier reasoning. Claude Sonnet 5, Sonnet 4.6, Sonnet 4.5 and Haiku 4.5 also receive context awareness tags that tell the model its remaining token budget after each tool call, while Opus 4.7 and later Opus, Sonnet 5.5, the Fable and Mythos 5.x models and Haiku 5.5 rely on task budgets, a beta feature, instead.


In the apps, the simplest controls are starting a fresh chat for a new task, keeping reference files in a project so retrieval can take over when they grow, and using memory or chat search to carry decisions from one conversation to the next. For Claude Code, the file-based memory in CLAUDE.md is described in our guide to Claude Code memory and persistent instructions.


··········


CLAUDE CONTEXT WINDOW, TOKEN LIMITS AND MEMORY IN NOVEMBER 2026.


The verified figures are consistent across Anthropic's documentation: 1M tokens of context and 128K tokens of output for every current model in the API, standard pricing across the whole window except for Haiku 5.5 above 100,000 tokens, and 200K tokens for the 4.5 generation. The apps are less uniform, with 500K tokens in chat for Fable 5, Opus 4.8, Opus 4.7, Opus 4.6 and Sonnet 4.6, separate figures for Claude Code and Cowork, and no published Free plan number.


Memory remains a separate layer from the context window. Inside a chat, automatic summarization extends long conversations for users with code execution enabled; across chats, topic-based memory, chat search and project retrieval decide what Claude can recall. The figures most likely to change next are the app-side windows for older models and the beta status of compaction and task budgets in the API, both of which Anthropic updates on its help center and platform documentation.


··········


FOLLOW US FOR MORE.


·····


DATA STUDIOS


·····


datastudios.org


bottom of page