Claude conversation too long? Why a 1M token chat still fills up, and 5 fixes

How do you fix a Claude conversation that is too long?

There are five fixes: turn on the setting that lets Claude summarize, hand off to a new chat, keep the work in a Project, split big files, and start long work on a 1M model. The first two take about two minutes:

1. Open Settings, then Capabilities, and turn on "Code execution and file creation" if it's off. Anthropic switches it on by default on Free, Pro and Max (on Team and Enterprise, an owner controls it under Organization settings). On paid plans, Claude's automatic summary of long chats only runs when it's on.
2. If the old chat still answers, paste the handoff prompt below and download the handoff.md file Claude saves.
3. Open a new chat on a 1M token model, inside a Project if the work lives in one. Attach handoff.md, or add it to the Project's knowledge so every chat there can read it.
4. Paste the opener below and read Claude's three line reply before you ask for anything new.

```text

We're moving this work to a new chat that can't see this one. Write a handoff document and save it as a Markdown file called handoff.md. Keep it under 800 words, with these headings:

1. Goal: what we're making or deciding, in two sentences.
2. Latest version: the current draft or code in full if it's under 300 words, otherwise its outline with every heading.
3. Decisions: one line each, with the reason.
4. Dead ends: what we tried or ruled out, and why, so it isn't suggested again.
5. Rules: every instruction I gave on tone, format, length, names and numbers.
6. Facts: every number, name, date and file name we rely on, exactly as written.
7. The next step, and any open questions.

Add nothing we didn't discuss. Mark anything you're unsure of with [check].

```

```text

handoff.md is a summary of an earlier chat that ran out of room. Read it, then reply in three lines only: the goal, the rules you'll follow, and the next step. Wait for my go before you start.

```

The file matters more than the summary. A handoff pasted as text sits in one chat. A handoff saved to project knowledge reaches every new chat in that Project, and you can edit it as the work moves on.

What if the old chat won't take another message? On a paid plan, open a new chat and ask Claude to search your past chats for the project by name. Anthropic's help center says you can prompt Claude "to search through your previous conversations", as long as Search and reference chats is on under Settings, Memory. Inside a Project, that search stays within the Project. On Free, copy the last good outputs by hand, or request your data from Settings, Privacy, Export data. The emailed link expires 24 hours after it arrives.

What does the Claude "too long" message mean?

Every version of the message means the same thing: this chat, plus what you're sending, is bigger than the model's context window. Waiting won't clear it. It's a size limit, not the usage cap that resets every five hours, which Claude usage limit reached covers.

The first message is the one Anthropic documents. Its error messages page suggests four ways out: send the content in smaller parts, summarize or extract the key sections, ask Claude to find the relevant parts first, or start a new conversation. It can appear on the very first message of a chat. A big enough upload fills the window by itself.

How long can a Claude conversation be?

A Claude conversation can hold 1M tokens on the newest models, 500K on the previous generation and 200K on the rest, on every paid plan, per Claude's context window page read on 11 October 2026. Anthropic says 200K is "about 500 pages of text or more". By that ratio, a 1M chat holds about 2,500 pages.

Upgrading from Pro to Max doesn't make a chat longer. The window belongs to the model, and Anthropic lists the same windows for every paid plan. Max buys more usage, not more room. Free gets no published figure. You also never get the full window for your own text. Anthropic's usage and length page says "A portion of the context window is always reserved for Claude's response", so the longest possible chat is a little shorter than the number in the table.

What fills a Claude chat so fast?

Files fill a Claude chat faster than typing does. Anthropic's PDF docs say each PDF page "typically uses 1,500 to 3,000 tokens" of text, depending on density. So one dense 100 page report can take 300,000 tokens before you ask a question. That's more than a whole 200K chat holds. A 1M chat fills at about 333 such pages.

PDFs of 100 pages or fewer carry pictures as well as text. Claude's file upload page says Claude analyzes "both text and visual elements" in them. Anthropic documents the token cost only for the API, where each such page is also priced as an image, so expect a short PDF to take more room than its text alone. Above 100 pages Claude reads text only, up to 1,000 pages. At 3,000 tokens a page, a 1,000 page PDF can reach 3M tokens, which no Claude chat window holds whole.

Images add up the same way. Anthropic's vision docs put a 1000 by 1000 pixel image at 1,296 tokens and a 4K screenshot at 4,784 tokens on recent models. Claude takes up to 20 images per message, so one message of large screenshots can cost about 96,000 tokens.

Four quieter things take room on every turn. Project knowledge is read as context in each chat in the Project until it grows big enough for Claude to switch to retrieval. Tools and connectors carry their own instructions, and Anthropic warns "Tools and connectors are token-intensive". Extended thinking and higher effort levels use more tokens per reply. And Claude rereads the whole history each time, so message 50 carries the 49 before it.

Split big files before they go in. Send the 20 pages you need, not the 300 page manual, or ask Claude to list the relevant sections first and attach only those. Turn off connectors and web search in chats that don't need them.

Does Claude compact long conversations automatically now?

Yes, on paid plans. Since 24 November 2025, when a chat nears its context limit, Claude summarizes earlier messages and keeps going, according to Anthropic's release notes. The usage and length limits page puts it this way: "When your conversation approaches the context window limit, Claude summarizes earlier messages to make room for new content." Your full history stays in the chat, and Claude can still refer back to it. While it works, you may see Claude "organizing its thoughts".

There are three catches. Anthropic's error messages page limits it to "users on paid plans with code execution enabled", so a Free chat, or one with Code execution off, can still hit the wall. Anthropic says rare edge cases, like a very large first message, can still run into the limit. And the summary costs you: "Longer conversations that trigger automatic context management consume more of your usage limit." That's why a long chat can drain a Pro session faster than several short ones.

A summary also loses detail. Claude Code's docs say the same about its own compaction: "detailed instructions from early in the conversation may be lost". Keep rules that must survive (tone, format, names) in project instructions, not in message 3 of a long chat.

ChatGPT behaves differently. It stops the thread and asks for a new chat, as ChatGPT conversation too long explains, which makes Claude's automatic summary the bigger practical difference between the two.

In Claude Code, compaction runs on its own and you can steer it. Run /compact with a focus, such as "/compact focus on the API changes", run /context to see what's using space, and put permanent rules in CLAUDE.md. If /compact fails with "Conversation too long", press Esc twice to step back a few messages, as the error says, and compact from there.

How do you keep a long Claude project from running out of room?

Put the reference material in a Claude Project and start one chat per task inside it. On paid plans, when project knowledge nears the context limit, Claude switches to retrieval and pulls in only the relevant parts. Anthropic's RAG for Projects page says that lets you "store up to 10x more content". Project files are capped at 30MB each, against 500MB for a file attached to a chat.

Keep project instructions short, since they ride along on every turn. Remove files you're done with. When a new chat needs something from an old one, ask Claude to search the Project's past chats instead of pasting history back in.

For a long single thread, start on a 1M model. A chat on a 200K model fills five times sooner than the same chat on Claude Opus or Claude Sonnet at 1M. Room isn't recall, though. What is a context window explains why models miss facts buried in the middle of long inputs, and the context window comparison lists windows model by model across vendors.

Whizi, which publishes this page, imports a Claude export in the browser and makes every message searchable, so an old thread you can no longer continue is still easy to mine. It drops thinking blocks, tool calls and attachments. It won't give a long thread more room: every Whizi model gets 40,000 input tokens per turn, and older history is trimmed silently. For one huge conversation, Claude's own 1M window with automatic summary is the better tool.