Signature experience · walkable window
Claude Fable 5.1, Claude Opus 5, and Claude Sonnet 5 each hold a 1M-token context window with 128K max output. Claude Haiku 4.5 holds 200K with 64K max output. History, the current input, tools, thinking blocks, and Claude’s own answer all count toward that window. Bigger is not automatically better. Walk the Corridor from Haiku’s 200K door into the 1M hall; markers show cache reads, compaction, what still fits, and what falls off.
ClaudeContext.com is an independent educational publication. It is not affiliated with, endorsed by, or sponsored by Anthropic.
Floor plan
You start on Haiku’s 200K side. Step the stations, then open the door into the 1M hall. The long stretch stays visible in shadow until you walk through.
Claude’s context window is working memory, not training data. Anthropic’s context-window docs describe it as accumulating across turns: prior messages stay unless you summarize, clear, or omit them. Tool definitions, tool results, thinking blocks where they must be preserved, and the answer Claude is about to write all compete for the same finite span.
Arrow keys step, space plays and pauses, and D opens Haiku’s door when the floor plan has focus. Cache-read dollars are official rates applied to the walker position; they are not a bill.
Stations
Landmarks at 0, 136K, 200K, 872K, and 1M are arithmetic on Anthropic’s published windows. The stations between them are editorial placements on the walk, labeled as such, so a crawler can tell a sourced number from a teaching mark.
Threshold
0 tokens · published landmark
Claude’s context window is working memory, not training data. Anthropic’s context-window docs describe it as accumulating across turns: prior messages stay unless you summarize, clear, or omit them. Tool definitions, tool results, thinking blocks where they must be preserved, and the answer Claude is about to write all compete for the same finite span.
What fits
48,000 tokens · editorial placement
What belongs in the hall is the information needed for the next call: purpose, instructions, a few examples, current sources, and the tools that will actually be used. Anthropic’s context-engineering essay treats context as finite because attention and recall degrade as irrelevant or conflicting tokens accumulate. The useful question is not “can this fit?” It is whether the token improves the next call enough to justify cost, latency, and distraction.
Editorial placement on the walk, not a published token budget.
Cache read
96,000 tokens · editorial placement
Prompt caching caches the full prefix. Cache reads cost 10 percent of base input on most Claude models. On Claude Fable 5.1 they cost 2.5 percent, which is $0.25 per million tokens against a $10 base input price. Mythos 5.1 uses that same 2.5 percent rate and is trusted-access only, so it stays off this selector. Caching does not shrink the window and does not make a bad prefix good. Selection first, cache second.
Editorial placement. Prices are official; this marker’s token coordinate is not a cache threshold.
Haiku output reserve
136,000 tokens · published landmark
Claude Haiku 4.5 lists a 200K-token window and 64K max output. The window includes the response, so a long answer competes with history. The hatched band from here to Haiku’s door is that 64K reserve: not a published “stop writing input” line, but the arithmetic of a 64K ceiling against a 200K wall.
Haiku’s door
200,000 tokens · published landmark
This is the published context-window limit for Claude Haiku 4.5. Claude Fable 5.1, Claude Opus 5, and Claude Sonnet 5 continue for another 800K tokens, each with 128K max output. Stepping through does not make the first 200K cheaper or more accurate. It only makes a longer hall available. If you stay on Haiku, everything past this door falls off.
What falls off
320,000 tokens · editorial placement · past Haiku’s door
On Haiku 4.5 this stretch does not exist. On a 1M model it exists, and it still costs attention. Long context is valuable when the task needs many documents, a large codebase, or a multi-step history. It is harmful when it becomes an excuse to keep stale instructions, duplicate examples, and unfiltered logs. The platform docs put it plainly: curating what is in context is as important as how much space is available.
Editorial placement past the published 200K door.
Compaction
520,000 tokens · editorial placement · past Haiku’s door
Compaction summarizes dialogue and continues from that compressed state. Tool-result clearing removes bulky artifacts that can be fetched again. Memory persists selected knowledge across sessions — a different product, covered by Claude Memories, not this corridor. Use compaction when the valuable state is the story of the conversation. Use clearing when old tool output is re-fetchable. Use memory when a fact must survive a fresh window. None of these is a substitute for a larger window, and a larger window is not a substitute for them.
Editorial placement. Compaction is a strategy, not a token trigger published here.
MCP resources
720,000 tokens · editorial placement · past Haiku’s door
MCP resources are URI-addressed readable context: files, schemas, documents, logs. They are not tools, which do work, and they are not prompts, which standardize a workflow. A resource inventory lets a client load the item needed for this step instead of pasting every relevant object into the prefix. Anthropic’s MCP connector is a Messages API path for remote server tools; it is not automatically the full protocol surface for resources.
Editorial placement on the long hall.
1M output reserve
872,000 tokens · published landmark · past Haiku’s door
Claude Fable 5.1, Claude Opus 5, and Claude Sonnet 5 each list 128K max output. As with Haiku, output tokens compete with input. A 1M window already full of history has no room to speak. Token counting should happen before the request, against the model that will actually handle it. Newer tokenizers can count the same text differently from earlier generations.
Far wall
1,000,000 tokens · published landmark · past Haiku’s door
This is the current generally available ceiling for Claude Fable 5.1, Claude Opus 5, and Claude Sonnet 5. It is not a target. Anthropic’s own writing treats a larger window as useful when the task needs it, and as a liability when it is used as a junk drawer. If you arrived here by stuffing, turn around: compact, retrieve, cache a smaller prefix, or expose the rest as MCP resources.
What this walk shows
What it does not show
Name the model first. The current generally available lineup is not one window. Claude Platform’s model overview lists Claude Fable 5.1, Claude Opus 5, and Claude Sonnet 5 at 1M tokens with 128K max output, and Claude Haiku 4.5 at 200K with 64K max output. The Corridor is that table turned into a floor.
Count what is already walking with you. Context-window docs treat the span as accumulating: history, the current input, tools, results, thinking blocks that must be preserved, and the answer about to be written. If a block has to come back on the next call, it is still in the hall.
Reserve the far end before you furnish the near end. Output is not extra. A 200K Haiku window with a 64K reply already spoken for has 136K of useful input; a 1M window with a 128K reply has 872K. The hatched bands on the floor are that arithmetic, not a hidden API flag.
Price a prefix only after you have chosen it. Prompt caching caches the full prefix. Pricing puts cache reads at 10 percent of base input on most models, and at 2.5 percent on Fable 5.1 — $0.25 per million tokens, the cut announced with Fable and Mythos 5.1. The walker multiplies those rates by its position so the cost of re-reading “this far” is visible. It will not save a bloated packet.
When the walk gets crowded, change the walk, not only the length of the building. Compaction keeps a long conversation usable. Tool-result clearing drops bulk you can fetch again. MCP resources address context by URI so you load the file this step needs. Durable facts that must survive a fresh window belong in memory, which is Claude Memories’ lane, not this one.
The existing tools remain the calculators. Use the context-window planner to allocate a real budget, the caching calculator to price a hit rate, and the compaction simulator to see what a summary keeps and drops. The Corridor is the map you walk before you open those.
ClaudeContext.com owns windows, compaction, prompts-as-context, MCP resources, and caching. It does not own the memory product. Topic-level edit and delete, chat and Cowork sharing a store, and who can turn memory on are documented on claudememories.com. Session transcripts as evidence belong on claudelogs.com. Claude Code practice belongs on claudeshipscode.com.
FAQ
Claude Fable 5.1, Claude Opus 5, and Claude Sonnet 5 each list a 1M-token context window with 128K max output. Claude Haiku 4.5 lists 200K with 64K max output. Checked against Claude Platform model docs on September 19, 2026.
The published window includes conversation history, the current input, tool definitions, tool-use records, tool results, relevant thinking blocks, and the response Claude generates. Output tokens compete with input for the same span.
No. Use the 1M hall when the task needs many documents, a large codebase, or a long history. Otherwise a smaller, curated window is usually cheaper, faster, and easier to recall from. Anthropic’s context-window docs say curating what is in context is as important as how much space is available.
200K is Claude Haiku 4.5’s published window. On Haiku it is a wall: anything past it falls off. On Fable 5.1, Opus 5, and Sonnet 5 it is a door into an 800K continuation of the same window. Stepping through changes which model’s limit you are walking, not whether the first 200K was well chosen.
Cache reads cost 10 percent of base input on most Claude models, including Opus 5 ($0.50 / MTok), Sonnet 5 ($0.20 / MTok), and Haiku 4.5 ($0.10 / MTok). Claude Fable 5.1 cache reads are 2.5 percent of base input, or $0.25 per million tokens. Mythos 5.1 uses that same 2.5 percent rate and is trusted-access only.
Compact when the valuable state is the story of a long conversation and raw history is crowding out the next decision. A larger window delays the problem; it does not decide what still deserves attention. Memory is a different lever for facts that must survive a fresh session, and is covered by Claude Memories rather than this site.
No. MCP resources expose readable context by URI so a client can load the item needed for this step. They are not tools and they are not a license to paste an entire inventory into the prompt.
No. ClaudeContext.com is an independent educational publication. It is not affiliated with, endorsed by, or sponsored by Anthropic.
Cite this page
Claude Context. “The Corridor: walk Claude’s 200K door into the 1M hall.” claudecontext.com, updated 2026-09-19. https://claudecontext.com/corridor/
ClaudeContext.com is an independent educational publication. It is not affiliated with, endorsed by, or sponsored by Anthropic.